MIRIX: Multi-Agent Memory System for LLM-Based Agents
% Yu Wang, Xi Chen, MIRIX AI
Abstract
Although memory capabilities of AI agents are gaining increasing attention, existing solutions remain fundamentally limited. Most rely on flat, narrowly scoped memory components, constraining their ability to personalize, abstract, and reliably recall user-specific information over time. To this end, we introduce MIRIX, a modular, multi-agent memory system that redefines the future of AI memory by solving the field’s most critical challenge: enabling language models to truly remember. Unlike prior approaches, MIRIX transcends text to embrace rich visual and multimodal experiences, making memory genuinely useful in real-world scenarios. MIRIX consists of six distinct, carefully structured memory types: Core, Episodic, Semantic, Procedural, Resource Memory, and Knowledge Vault, coupled with a multi-agent framework that dynamically controls and coordinates updates and retrieval. This design enables agents to persist, reason over, and accurately retrieve diverse, long-term user data at scale. We validate MIRIX in two demanding settings. First, on ScreenshotVQA, a challenging multimodal benchmark comprising nearly 20,000 high-resolution computer screenshots per sequence, requiring deep contextual understanding and where no existing memory systems can be applied, MIRIX achieves 35% higher accuracy than the RAG baseline while reducing storage requirements by 99.9%. Second, on LOCOMO, a long-form conversation benchmark with single-modal textual input, MIRIX attains state-of-the-art performance of 85.4%, far surpassing existing baselines. These results show that MIRIX sets a new performance standard for memory-augmented LLM Agents. To allow users to experience our memory system, we provide a packaged application powered by \ours. It monitors the screen in real time, builds a personalized memory base, and offers intuitive visualization and secure local storage to ensure privacy.
Multi-Agent Systems
Yu Wang, Xi Chen MIRIX AI
yuw164@ucsd.edu xc13@stern.nyu.edu https://mirix.io/
Chat Interface
Although memory capabilities of AI agents are gaining increasing attention, existing solutions remain fundamentally limited. Most rely on flat, narrowly scoped memory components, constraining their ability to personalize, abstract, and reliably recall user-specific information over time. To this end, we introduce MIRIX, a modular, multi-agent memory system that redefines the future of AI memory by solving the field's most critical challenge: enabling language models to truly remember. Unlike prior approaches, MIRIX transcends text to embrace rich visual and multimodal experiences, making memory genuinely useful in real-world scenarios. MIRIX consists of six distinct, carefully structured memory types: Core, Episodic, Semantic, Procedural, Resource Memory, and Knowledge Vault, coupled with a multi-agent framework that dynamically controls and coordinates updates and retrieval. This design enables agents to persist, reason over, and accurately retrieve diverse, long-term user data at scale. We validate MIRIX in two demanding settings. First, on ScreenshotVQA, a challenging multimodal benchmark comprising nearly 20,000 high-resolution computer screenshots per sequence, requiring deep contextual understanding and where no existing memory systems can be applied, MIRIX achieves 35% higher accuracy than the RAG baseline while reducing storage requirements by 99.9%. Second, on LOCOMO, a long-form conversation benchmark with single-modal textual input, MIRIX attains state-of-the-art performance of 85.4%, far surpassing existing baselines. These results show that MIRIX sets a new performance standard for memory-augmented LLM Agents. To allow users to experience our memory system, we provide a packaged application powered by MIRIX. It monitors the screen in real time, builds a personalized memory base, and offers intuitive visualization and secure local storage to ensure privacy.
Introduction
Recent advancements in large language model (LLM) agents have focused primarily on improving their capabilities in complex task execution-ranging from code debugging and repository management to autonomous web browsing. While these functionalities are crucial, another foundational yet underexplored dimension is memory: the ability of agents to persist, retrieve, and utilize past user-specific information over time. Human cognition relies heavily on memory-recalling conversations, recognizing patterns, and adapting behavior based on prior experience. Analogously, memory mechanisms in LLM agents are essential for delivering consistent, personalized interactions, learning from feedback, and avoiding repetitive queries. However, most LLM-based personal assistants remain stateless beyond their current prompt window, retaining no lasting memory unless context is explicitly re-provided. This limitation hinders their long-term usability, especially in real-world settings where users expect assistants to evolve, recall, and personalize over time [29].
Figure 1: The six memory components of MIRIX, each providing specialized functionality.
To address this, a range of memory-augmented systems have been proposed. One common approach is the use of knowledge graphs, as seen in systems like Zep [24] and Cognee [20]. These frameworks are well-suited for representing structured relationships between entities but struggle to model sequential events, emotional states, full-length documents, or multi-modal inputs such as images. Another approach involves flattened memory architectures that store and retrieve textual chunks using vector databases. Examples include Letta [22], Mem0 [5], and ChatGPT's memory system. Letta divides its memory into components-recall memory for conversation history, core memory for preferences, and archival memory for long documents-while ChatGPT focuses primarily on core and recall memories. Mem0 adopts a memory system that contains flattened facts distilled from the user inputs, which serves similar roles as Letta's archival memory while being more distilled. While prevalent, these memory systems face several challenges: (1) Lack of compositional memory structure: Most approaches store all historical data in a single flat store without routing into specialized memory types (e.g., procedural, episodic, semantic), making retrieval inefficient and less accurate. (2) Poor multi-modal support: Text-centric memory mechanisms fail when the majority of the input is nonverbal (e.g., images, interface layouts, maps). (3) Scalability and abstraction: Storing raw inputs, especially images, leads to prohibitive memory requirements, with no effective abstraction layer to summarize and retain only salient information.
To address the limitations of existing memory systems, we argue that effective Routing and Retrieving are the key capabilities a memory-augmented agent must possess. Most current systems focus primarily on Short-Term and Long-Term Memory [22, 32, 25], with some incorporating Mid-Term Memory [12]. In contrast, we draw inspiration from works that explore more specialized memory types, including episodic memory [23, 18, 1], semantic memory [15], and procedural memory [33]. Building on these foundations, we propose a more comprehensive architecture consisting of six memory components: Core Memory , Episodic Memory , Semantic Memory , Procedural Memory , Resource Memory , and the Knowledge Vault (As shown in Figure 1). Episodic Memory stores user-specific events and experiences; Semantic Memory captures concepts and named entities (e.g., the meaning of a new phrase or an understanding of a person); Procedural Memory records step-bystep instructions for performing tasks. Resource Memory is designed to store documents, files, and other media shared by the user. Knowledge Vault holds critical verbatim information that must be preserved exactly, such as addresses, phone numbers, email accounts, and other sensitive facts.
Each memory component is internally organized using a hierarchical structure. For example, Episodic Memory includes fields such as summary and details , while Semantic Memory organizes information by name and description . Managing this structured and heterogeneous memory is challenging for a single agent. Therefore, we adopt a multi-agent architecture: six Memory Managers for six memory components, and a Meta Memory Manager responsible for task routing. While this memory system can be plugged and connected with other existing agents, we build an extra Chat Agent to demonstrate how we can interact with an agent that has access to our memories. This memory system, which we call MIRIX , is modular and designed as a comprehensive and full memory system for LLM-based agents. Moreover, when interacting with the Chat Agent, we propose an Active Retrieval mechanism, where the agent is required to generate a topic before answering the question or executing the next step, and the retrieved information is inputted into the model in the system prompt. Meanwhile, we design multiple retrieval tools so that the agent can choose appropriate ones in response to different situations.
We evaluate MIRIX in two experimental settings. First, we introduce a challenging benchmark that requires extracting information and building memory from multimodal input. To this end, we
collect between 5,000 and 20,000 high-resolution screenshots spanning one month of computer usage from three PhD students and construct evaluation questions grounded in their visual activity history. This setting demands sophisticated memory modeling and exceeds the capacity of existing long-context models. For example, each screenshot ranges from 2K to 4K resolution depending on the user's monitor, which limits Gemini to processing no more than 500 images at full resolution or approximately 3,600 images when resized to 256×256 pixels. Second, following Mem0 [5], we assess performance on the LOCOMO dataset [19], which comprises long-form, multi-turn conversations. While each conversation is relatively short (around 26,000 tokens on average), we constrain the Chat Agent to answer questions using only retrieved memories, without access to the original conversation transcripts. This setup enables us to evaluate whether our system can effectively distill and route essential information into memory. In terms of results, because no existing memory systems can handle such a large volume of multimodal input, we compare against retrieval-augmented generation (RAG) baselines and long-context baselines. MIRIX achieves a 35% improvement over RAG baselines while reducing storage requirements by 99.9%, and a 410% improvement over longcontext baselines with a 93.3% reduction in storage. On LOCOMO, MIRIX reaches state-of-the-art performance with an overall accuracy of 85.38%, outperforming the best existing method by 8.0% and approaching the upper bound set by long-context models. To make our memory system more accessible, we also developed a personal assistant powered by MIRIX. With the user's permission to capture screen content, the assistant continuously builds a memory from screenshots and can answer any questions related to this accumulated information. This represents our effort to bring advanced memory capabilities to everyone, enabling users to experience the benefits of our system firsthand.
Our contributions are summarized as follows:
· We analyze the limitations of existing memory architectures for LLM-based agents and propose a novel memory system composed of six specialized components and eight different agents. · We introduce a new benchmark in which the agent must interpret a large collection of screenshots to build an understanding of user behavior-posing a significant challenge to the memory. · Through experiments on our new benchmark and existing benchmark LOCOMO, we demonstrate that MIRIX significantly outperform existing memory systems. · To make our approach accessible, we build and release a personal assistant application powered by MIRIX, enabling users to experience advanced memory capabilities in real-world scenarios.
Application & Use Cases
ours Application
To demonstrate the full functionality of our memory system, we developed a cross-platform application using React-Electron for the frontend and Uvicorn as the backend server. Moreover, we release the obtained file for direct installation. Within the application, users can activate screen monitoring by selecting the ScreenShots tab. This enables the agent to observe the user's screen activity and dynamically update its memory, gradually building a contextual understanding of the user over time. With the obtained memory, the agent can answer questions related to the memories.
Memory Updates The application captures a screenshot every 1.5 seconds. To reduce redundancy, images that are visually similar to previously captured ones are discarded. Once 20 unique screenshots are collected, the memory update process is triggered-typically around every 60 seconds. During this process, relevant information is extracted and incorporated into the system's memory components. To reduce latency in processing user screenshots, we adopt a streaming upload strategy. Instead of batching and sending 20 images at once, we upload each screenshot immediately upon receiving it from the front-end. By leveraging the Gemini API-which supports loading images via Google Cloud URLs-we can efficiently transmit visual data without waiting for the full batch to accumulate. This approach significantly reduces end-to-end latency from approximately 50 seconds (as observed when using GPT-4 with direct image upload) to under 5 seconds using Gemini.
Chat Interface The chat interface allows users to interact with the agent, which has full access to its accumulated memories. As illustrated in Figure 2, users can query the assistant about past activities, and the agent can respond based on its memory content, enabling more informed and personalized interactions.
Figure 2: Chat Window
Memory Visualization After observing the screen for a sufficient period, the agent organizes its knowledge into structured memory components. An example of Semantic Memory (organized into a tree structure) is shown in Figure 3. We also provide a list view of the memories. An example of Procedural Memory is shown in Figure 4.
Memory Updates
We design a modular memory architecture consisting of six distinct components: Core Memory , Episodic Memory , Semantic Memory , Procedural Memory , Resource Memory , and the Knowledge Vault . Each component is structurally and functionally tailored to capture different aspects of user interaction and world knowledge, enabling the agent to retrieve, reason, and act effectively across time and tasks.
Core Memory Core Memory stores high-priority, persistent information that should always remain visible to the agent when engaging with the user. Inspired by the design of MemGPT [22], this memory is divided into two primary blocks: persona and human . The persona block encodes the identity, tone, or behavior profile of the agent, while the human block stores enduring facts about the user such as their name, preferences, and self-identifying attributes (e.g., 'User's name is David', 'User enjoys Japanese cuisine'). When the memory size exceeds 90% of capacity, the system triggers a controlled rewrite process to maintain compactness without losing critical information.
Episodic Memory Episodic Memory captures time-stamped events and temporally grounded interactions that reflect the user's behavior, experiences, or activities. It functions as a structured log or calendar, enabling the agent to reason about user routines, recency, and context-aware followups [17, 23, 18]. Each entry is defined by the following fields: event_type (e.g., user_message , inferred_result , system_notification ), summary (a concise natural language description of the event), details (extended contextual information, including dialog excerpts or inferred states), actor (the origin of the event, either user or assistant ), and timestamp (e.g., '2025-03-05 10:15'). This structure allows the agent to temporally index memory and track change over time, such as identifying ongoing tasks or following up on pending actions.
Figure 5: Demonstration of Active Retrieval.
Semantic Memory Semantic Memory maintains abstract knowledge and factual information that is independent of specific times or events. This component serves as a knowledge base for general concepts, entities, and relationships - whether about the world or the user's social graph. For example, it may store entries like 'Harry Potter is written by J.K. Rowling' or 'John is a friend of the user who enjoys jogging and lives in San Francisco.' Each entry includes a name (the concept or entity identifier), summary (a concise definition or relationship statement), details (expanded background or contextual explanation), and source (e.g., user_provided , Wikipedia , or inferred from conversation). Unlike episodic memory, semantic entries are intended to persist unless conceptually overwritten and support reasoning over social, geographic, or commonsense knowledge.
Procedural Memory Procedural Memory stores structured, goal-directed processes such as how-to guides, operational workflows, and interactive scripts. These are neither time-sensitive (episodic) nor abstract facts (semantic) but instead represent actionable knowledge that can be invoked to assist the user with complex tasks. Typical examples include 'how to file a travel reimbursement form,' 'steps to set up a Zoom meeting,' or 'how to book a restaurant via OpenTable.' Each entry includes an entry_type ( workflow , guide , or script ), a description of the goal or function, and steps expressed as a list of instructions (optionally in JSON or structured format). This memory component supports instructional planning, automation, and decomposition of user goals into sub-tasks.
Resource Memory Resource Memory handles full or partial documents, transcripts, or multi-modal files that the user is actively engaged with but do not fit into other memory categories. For instance, if the user is reading a friend's detailed picnic plan or a project proposal document, the agent can store and retrieve that information from Resource Memory. This component enables context continuity in long-running tasks. Each entry includes a title (resource name), summary (brief overview and context), resource_type (e.g., doc , markdown , pdf_text , image , voice_transcript ), and the full or excerpted content . This design enables the agent to reference previously seen material, quote from documents, or search within them to aid the user's workflow.
Knowledge Vault The Knowledge Vault serves as a secure repository for verbatim and sensitive information such as credentials, addresses, contact information, and API keys. These entries are not typically relevant to conversation-level reasoning but are crucial for performing authenticated tasks or storing long-term identifiers. Each entry includes an entry_type (e.g., credential , bookmark , contact_info , api_key ), a source (e.g., user_provided , github ), a sensitivity level ( low , medium , high ), and the actual secret_value . Entries with high sensitivity are protected via access control and excluded from casual retrieval to prevent misuse or leakage.
Chat Interface
Although memory capabilities of AI agents are gaining increasing attention, existing solutions remain fundamentally limited. Most rely on flat, narrowly scoped memory components, constraining their ability to personalize, abstract, and reliably recall user-specific information over time. To this end, we introduce MIRIX, a modular, multi-agent memory system that redefines the future of AI memory by solving the field's most critical challenge: enabling language models to truly remember. Unlike prior approaches, MIRIX transcends text to embrace rich visual and multimodal experiences, making memory genuinely useful in real-world scenarios. MIRIX consists of six distinct, carefully structured memory types: Core, Episodic, Semantic, Procedural, Resource Memory, and Knowledge Vault, coupled with a multi-agent framework that dynamically controls and coordinates updates and retrieval. This design enables agents to persist, reason over, and accurately retrieve diverse, long-term user data at scale. We validate MIRIX in two demanding settings. First, on ScreenshotVQA, a challenging multimodal benchmark comprising nearly 20,000 high-resolution computer screenshots per sequence, requiring deep contextual understanding and where no existing memory systems can be applied, MIRIX achieves 35% higher accuracy than the RAG baseline while reducing storage requirements by 99.9%. Second, on LOCOMO, a long-form conversation benchmark with single-modal textual input, MIRIX attains state-of-the-art performance of 85.4%, far surpassing existing baselines. These results show that MIRIX sets a new performance standard for memory-augmented LLM Agents. To allow users to experience our memory system, we provide a packaged application powered by MIRIX. It monitors the screen in real time, builds a personalized memory base, and offers intuitive visualization and secure local storage to ensure privacy.
Memory Visualization
To demonstrate the full functionality of our memory system, we developed a cross-platform application using React-Electron for the frontend and Uvicorn as the backend server. Moreover, we release the obtained file for direct installation. Within the application, users can activate screen monitoring by selecting the ScreenShots tab. This enables the agent to observe the user's screen activity and dynamically update its memory, gradually building a contextual understanding of the user over time. With the obtained memory, the agent can answer questions related to the memories.
Memory Updates The application captures a screenshot every 1.5 seconds. To reduce redundancy, images that are visually similar to previously captured ones are discarded. Once 20 unique screenshots are collected, the memory update process is triggered-typically around every 60 seconds. During this process, relevant information is extracted and incorporated into the system's memory components. To reduce latency in processing user screenshots, we adopt a streaming upload strategy. Instead of batching and sending 20 images at once, we upload each screenshot immediately upon receiving it from the front-end. By leveraging the Gemini API-which supports loading images via Google Cloud URLs-we can efficiently transmit visual data without waiting for the full batch to accumulate. This approach significantly reduces end-to-end latency from approximately 50 seconds (as observed when using GPT-4 with direct image upload) to under 5 seconds using Gemini.
Chat Interface The chat interface allows users to interact with the agent, which has full access to its accumulated memories. As illustrated in Figure 2, users can query the assistant about past activities, and the agent can respond based on its memory content, enabling more informed and personalized interactions.
Figure 2: Chat Window
Memory Visualization After observing the screen for a sufficient period, the agent organizes its knowledge into structured memory components. An example of Semantic Memory (organized into a tree structure) is shown in Figure 3. We also provide a list view of the memories. An example of Procedural Memory is shown in Figure 4.
Memory System for Wearable Devices
The wearable device market has seen rapid growth in recent years, driven by the increasing demand for intelligent, always-available personal assistants. Products like AI-powered glasses (e.g., Meta RayBan, XREAL Air) and AI pins (e.g., Humane, Rabbit R1) aim to integrate seamless interaction into daily life through voice commands, visual capture, and real-time feedback. However, these devices often lack a long-term memory component that allows them to evolve with the user-retaining useful information over time, adapting to personal routines, and referencing past interactions in context-aware ways.
Our memory system is well-suited for integration into such wearable devices. By continuously collecting and processing data streams such as audio, visual scenes, and user queries, our system enables real-time memory formation. For instance, AI glasses equipped with our memory systems can automatically summarize meetings, remember frequently visited places, recognize recurring visual patterns, and recall previous conversations or tasks. With MIRIX, it can also evolve with the user and build a memory that is specifically made for the user. Moreover, the system's modular memory architecture-including procedural, episodic, semantic, and resource memory-aligns naturally with the needs of lightweight, on-the-go devices. Procedural memory enables the assistant to learn user habits (e.g., daily routes, meeting structures), while semantic memory stores general knowledge about the user's preferences, environment, and routines. Episodic memory captures time-stamped, situational experiences, and can be queried for recalling specific events (e.g., 'What did I see at the conference last week?'). Semantic Memory, on the other hand, can help organize the clients someone has seen over the past week and list them in a tree structure with details of their discussions.
Given the constraints of wearable hardware (limited compute and storage), our design also supports hybrid on-device/cloud memory management. Critical information in Knowledge Vault can be stored locally while large-scale memories such as Resource Memory can be offloaded and retrieved from the cloud on demand.
Figure 3: Tree Structure of an Example Semantic Memory. In this example, the user's Semantic Memory, which stores new concepts and their relationships, is organized hierarchically into multiple categories, such as Social Network and Favorites . Within Favorites , the memory is further divided into more specific classes, including Sports , Pets and Music .
Figure 4: List View of an Example Procedural Memory.
In summary, our memory system serves as a cognitive backbone for wearable AI agents-enabling personalization, continuity, and intelligence at the edge. As the wearable market matures, embedding persistent, structured memory will be a key differentiator for next-generation AI assistants.
Agent Memory Marketplace
We envision a future where personal memory-collected and structured through AI agents-becomes a new digital asset class. In the AI era, memory is no longer just a passive log of past events, but an active, evolving knowledge base that can be shared, personalized, and monetized. The Agent Memory Marketplace is our proposal for a decentralized ecosystem where memory is exchanged, reused, and built collaboratively through AI agents.
We begin with a core belief: human memory will become the most valuable and irreplaceable asset in the age of AI . Unlike static data, memory encompasses lived experiences, subjective context,
preferences, and interactions-making it deeply personal, yet highly reusable by intelligent systems. the marketplace is structured into three key layers:
- AI Agents Infrastructure . Our technology provides infrastructure for lifetime intelligent, interactive agents. Examples include (1) Personal AI Assistants & Companions: Tailored agents that continuously learn and evolve with the user. (2) AI Wearables: AI agents embedded in decentralized physical infrastructure (e.g., smart glasses, AI pins) that extend memory capture to the real world. (3) Multi-Agent Systems: Collaborative agents with shared memory access, enabling coordination and collective intelligence.
- Privacy-Preserving Memory Infrastructure To support trust and adoption, we aim to build in a robust privacy layer: (1) Encryption Layer: All memories are stored using end-to-end encryption. (2) Privacy Control: Fine-grained permissions allow users to choose which parts of their memory to share, trade, or restrict. (3) Decentralized Storage: Memories are stored in a distributed, censorship-resistant infrastructure.
- Memory Marketplace and Social Function : A peer-to-peer ecosystem for sharing, aggregating, and trading memories encoded in AI agents. Use cases include: (1) Memory Social/Trading: Tokenized exchange of memories (e.g., productivity hacks, niche workflows, or life advice). (2) Expert Communities: Collective memory-building for domain-specific expertise such as finance, education, or pet care. (3) Fan Economy and Dating Applications: Users can subscribe to AI personas based on influencer or celebrity memories, allowing fans to interact with memory-rich digital replicas of their favorite personalities. This technology also enables the creation of AI clones for accelerated dating and matching processes, where users can engage with AI agents in interactions before meeting in person.
In summary, we envision a future where personal memory transcends its traditional role to become an active, valuable digital asset. By combining lifelong AI agents, privacy-preserving infrastructure, and a decentralized marketplace, we aim to create an ecosystem where memories can be securely captured, meaningfully shared, and collectively advanced. This approach not only empowers individuals to benefit from their own experiences but also unlocks new possibilities for collaboration, personalization, and economic value in the AI era.
Methodology
Memory Components
We design a modular memory architecture consisting of six distinct components: Core Memory , Episodic Memory , Semantic Memory , Procedural Memory , Resource Memory , and the Knowledge Vault . Each component is structurally and functionally tailored to capture different aspects of user interaction and world knowledge, enabling the agent to retrieve, reason, and act effectively across time and tasks.
Core Memory Core Memory stores high-priority, persistent information that should always remain visible to the agent when engaging with the user. Inspired by the design of MemGPT [22], this memory is divided into two primary blocks: persona and human . The persona block encodes the identity, tone, or behavior profile of the agent, while the human block stores enduring facts about the user such as their name, preferences, and self-identifying attributes (e.g., 'User's name is David', 'User enjoys Japanese cuisine'). When the memory size exceeds 90% of capacity, the system triggers a controlled rewrite process to maintain compactness without losing critical information.
Episodic Memory Episodic Memory captures time-stamped events and temporally grounded interactions that reflect the user's behavior, experiences, or activities. It functions as a structured log or calendar, enabling the agent to reason about user routines, recency, and context-aware followups [17, 23, 18]. Each entry is defined by the following fields: event_type (e.g., user_message , inferred_result , system_notification ), summary (a concise natural language description of the event), details (extended contextual information, including dialog excerpts or inferred states), actor (the origin of the event, either user or assistant ), and timestamp (e.g., '2025-03-05 10:15'). This structure allows the agent to temporally index memory and track change over time, such as identifying ongoing tasks or following up on pending actions.
Figure 5: Demonstration of Active Retrieval.
Semantic Memory Semantic Memory maintains abstract knowledge and factual information that is independent of specific times or events. This component serves as a knowledge base for general concepts, entities, and relationships - whether about the world or the user's social graph. For example, it may store entries like 'Harry Potter is written by J.K. Rowling' or 'John is a friend of the user who enjoys jogging and lives in San Francisco.' Each entry includes a name (the concept or entity identifier), summary (a concise definition or relationship statement), details (expanded background or contextual explanation), and source (e.g., user_provided , Wikipedia , or inferred from conversation). Unlike episodic memory, semantic entries are intended to persist unless conceptually overwritten and support reasoning over social, geographic, or commonsense knowledge.
Procedural Memory Procedural Memory stores structured, goal-directed processes such as how-to guides, operational workflows, and interactive scripts. These are neither time-sensitive (episodic) nor abstract facts (semantic) but instead represent actionable knowledge that can be invoked to assist the user with complex tasks. Typical examples include 'how to file a travel reimbursement form,' 'steps to set up a Zoom meeting,' or 'how to book a restaurant via OpenTable.' Each entry includes an entry_type ( workflow , guide , or script ), a description of the goal or function, and steps expressed as a list of instructions (optionally in JSON or structured format). This memory component supports instructional planning, automation, and decomposition of user goals into sub-tasks.
Resource Memory Resource Memory handles full or partial documents, transcripts, or multi-modal files that the user is actively engaged with but do not fit into other memory categories. For instance, if the user is reading a friend's detailed picnic plan or a project proposal document, the agent can store and retrieve that information from Resource Memory. This component enables context continuity in long-running tasks. Each entry includes a title (resource name), summary (brief overview and context), resource_type (e.g., doc , markdown , pdf_text , image , voice_transcript ), and the full or excerpted content . This design enables the agent to reference previously seen material, quote from documents, or search within them to aid the user's workflow.
Knowledge Vault The Knowledge Vault serves as a secure repository for verbatim and sensitive information such as credentials, addresses, contact information, and API keys. These entries are not typically relevant to conversation-level reasoning but are crucial for performing authenticated tasks or storing long-term identifiers. Each entry includes an entry_type (e.g., credential , bookmark , contact_info , api_key ), a source (e.g., user_provided , github ), a sensitivity level ( low , medium , high ), and the actual secret_value . Entries with high sensitivity are protected via access control and excluded from casual retrieval to prevent misuse or leakage.
Core Memory
We envision a future where personal memory-collected and structured through AI agents-becomes a new digital asset class. In the AI era, memory is no longer just a passive log of past events, but an active, evolving knowledge base that can be shared, personalized, and monetized. The Agent Memory Marketplace is our proposal for a decentralized ecosystem where memory is exchanged, reused, and built collaboratively through AI agents.
We begin with a core belief: human memory will become the most valuable and irreplaceable asset in the age of AI . Unlike static data, memory encompasses lived experiences, subjective context,
preferences, and interactions-making it deeply personal, yet highly reusable by intelligent systems. the marketplace is structured into three key layers:
- AI Agents Infrastructure . Our technology provides infrastructure for lifetime intelligent, interactive agents. Examples include (1) Personal AI Assistants & Companions: Tailored agents that continuously learn and evolve with the user. (2) AI Wearables: AI agents embedded in decentralized physical infrastructure (e.g., smart glasses, AI pins) that extend memory capture to the real world. (3) Multi-Agent Systems: Collaborative agents with shared memory access, enabling coordination and collective intelligence.
- Privacy-Preserving Memory Infrastructure To support trust and adoption, we aim to build in a robust privacy layer: (1) Encryption Layer: All memories are stored using end-to-end encryption. (2) Privacy Control: Fine-grained permissions allow users to choose which parts of their memory to share, trade, or restrict. (3) Decentralized Storage: Memories are stored in a distributed, censorship-resistant infrastructure.
- Memory Marketplace and Social Function : A peer-to-peer ecosystem for sharing, aggregating, and trading memories encoded in AI agents. Use cases include: (1) Memory Social/Trading: Tokenized exchange of memories (e.g., productivity hacks, niche workflows, or life advice). (2) Expert Communities: Collective memory-building for domain-specific expertise such as finance, education, or pet care. (3) Fan Economy and Dating Applications: Users can subscribe to AI personas based on influencer or celebrity memories, allowing fans to interact with memory-rich digital replicas of their favorite personalities. This technology also enables the creation of AI clones for accelerated dating and matching processes, where users can engage with AI agents in interactions before meeting in person.
In summary, we envision a future where personal memory transcends its traditional role to become an active, valuable digital asset. By combining lifelong AI agents, privacy-preserving infrastructure, and a decentralized marketplace, we aim to create an ecosystem where memories can be securely captured, meaningfully shared, and collectively advanced. This approach not only empowers individuals to benefit from their own experiences but also unlocks new possibilities for collaboration, personalization, and economic value in the AI era.
Episodic Memory
Semantic Memory
We envision a future where personal memory-collected and structured through AI agents-becomes a new digital asset class. In the AI era, memory is no longer just a passive log of past events, but an active, evolving knowledge base that can be shared, personalized, and monetized. The Agent Memory Marketplace is our proposal for a decentralized ecosystem where memory is exchanged, reused, and built collaboratively through AI agents.
We begin with a core belief: human memory will become the most valuable and irreplaceable asset in the age of AI . Unlike static data, memory encompasses lived experiences, subjective context,
preferences, and interactions-making it deeply personal, yet highly reusable by intelligent systems. the marketplace is structured into three key layers:
- AI Agents Infrastructure . Our technology provides infrastructure for lifetime intelligent, interactive agents. Examples include (1) Personal AI Assistants & Companions: Tailored agents that continuously learn and evolve with the user. (2) AI Wearables: AI agents embedded in decentralized physical infrastructure (e.g., smart glasses, AI pins) that extend memory capture to the real world. (3) Multi-Agent Systems: Collaborative agents with shared memory access, enabling coordination and collective intelligence.
- Privacy-Preserving Memory Infrastructure To support trust and adoption, we aim to build in a robust privacy layer: (1) Encryption Layer: All memories are stored using end-to-end encryption. (2) Privacy Control: Fine-grained permissions allow users to choose which parts of their memory to share, trade, or restrict. (3) Decentralized Storage: Memories are stored in a distributed, censorship-resistant infrastructure.
- Memory Marketplace and Social Function : A peer-to-peer ecosystem for sharing, aggregating, and trading memories encoded in AI agents. Use cases include: (1) Memory Social/Trading: Tokenized exchange of memories (e.g., productivity hacks, niche workflows, or life advice). (2) Expert Communities: Collective memory-building for domain-specific expertise such as finance, education, or pet care. (3) Fan Economy and Dating Applications: Users can subscribe to AI personas based on influencer or celebrity memories, allowing fans to interact with memory-rich digital replicas of their favorite personalities. This technology also enables the creation of AI clones for accelerated dating and matching processes, where users can engage with AI agents in interactions before meeting in person.
In summary, we envision a future where personal memory transcends its traditional role to become an active, valuable digital asset. By combining lifelong AI agents, privacy-preserving infrastructure, and a decentralized marketplace, we aim to create an ecosystem where memories can be securely captured, meaningfully shared, and collectively advanced. This approach not only empowers individuals to benefit from their own experiences but also unlocks new possibilities for collaboration, personalization, and economic value in the AI era.
Procedural Memory
Resource Memory
Knowledge Vault
Active Retrieval and Retrieval Design
In many memory-augmented systems (e.g., Mem0 [5], MemGPT [22]), memory retrieval must be explicitly triggered. Otherwise, the language model often defaults to its parametric knowledge, which may be outdated or incorrect. For example, suppose the user previously said, 'The CEO of Twitter is Linda Yaccarino,' and this information was saved in memory. A few days later, when this message is no longer present in the conversation history, the user might ask, 'Who is the CEO of Twitter?' In
Figure 6: The Workflow of Memory Update.
this case, the language model may rely on outdated knowledge and incorrectly respond with 'Elon Musk.' While explicitly instructing the model to 'search your memory' can mitigate such errors, doing so repeatedly is impractical in natural conversations.
To address this, we propose an Active Retrieval mechanism. As illustrated in Figure 5, the system operates in two stages: first, the agent generates a current topic based on the input context; second, this topic is used to retrieve relevant memories from each memory component. Retrieved results are then injected into the system prompt. For instance, given the query 'Who is the CEO of Twitter?', the agent may infer the topic 'CEO of Twitter', which is then used to retrieve the top-10 most relevant entries from each of the six memory components. Retrieved content is tagged according to its source, such as <episodic_memory> ... </episodic_memory> , ensuring the model is aware of both the content and its origin. This automatic retrieval pipeline eliminates the need for explicit user prompts to trigger memory access and ensures the model can incorporate up-to-date, personalized, or contextual information during response generation.
In addition to Active Retrieval, we support multiple retrieval functions, including embedding_match , bm25_match , and string_match . We are actively expanding this set with more diverse and specialized retrieval strategies, ensuring that each method is well-differentiated so the agent can choose the most appropriate one to invoke based on context.
Multi-Agent Workflow
To manage the dynamic and heterogeneous nature of user interactions, we adopt a modular multiagent architecture. This system orchestrates input processing, memory updating, and information retrieval across six distinct memory components through a coordinated and efficient workflow. The overall system is governed by a central Meta Memory Manager and a set of specialized Memory Managers , each responsible for maintaining one memory type.
Memory Update Workflow As illustrated in Figure 7, when new input is received from the user, the system first automatically performs a search over the memory base. The retrieved information, together with the user input, is passed to the Meta Memory Manager . The Meta Memory Manager then analyzes the content and determines which memory components are relevant, routing the input to the corresponding Memory Managers . These Memory Managers update their respective memories in parallel while ensuring that redundant information is avoided within each memory type. After
Figure 7: The Workflow of Responding to the User's Query.
completing the updates, they report back to the Meta Memory Manager , which finally sends an acknowledgment confirming that the memory update process is complete.
Conversational Retrieval Workflow For interactive dialogues, the Chat Agent manages natural language communication with the user. To ground its responses in prior knowledge, it first performs an automatic search over the memory base upon receiving a user query. This initial search is a coarse retrieval spanning all six memory components and returns high-level summaries rather than detailed content. The Chat Agent then analyzes the query to determine which memory components warrant more targeted searches and selects appropriate retrieval methods accordingly. After obtaining the relevant results, it consolidates the information and synthesizes the final response. Moreover, if the user's query involves updating memory-for example, providing new facts or corrections-the Chat Agent can interact directly with the corresponding Memory Managers to apply precise updates to specific memory components.
Memory Update Workflow
To manage the dynamic and heterogeneous nature of user interactions, we adopt a modular multiagent architecture. This system orchestrates input processing, memory updating, and information retrieval across six distinct memory components through a coordinated and efficient workflow. The overall system is governed by a central Meta Memory Manager and a set of specialized Memory Managers , each responsible for maintaining one memory type.
Memory Update Workflow As illustrated in Figure 7, when new input is received from the user, the system first automatically performs a search over the memory base. The retrieved information, together with the user input, is passed to the Meta Memory Manager . The Meta Memory Manager then analyzes the content and determines which memory components are relevant, routing the input to the corresponding Memory Managers . These Memory Managers update their respective memories in parallel while ensuring that redundant information is avoided within each memory type. After
Figure 7: The Workflow of Responding to the User's Query.
completing the updates, they report back to the Meta Memory Manager , which finally sends an acknowledgment confirming that the memory update process is complete.
Conversational Retrieval Workflow For interactive dialogues, the Chat Agent manages natural language communication with the user. To ground its responses in prior knowledge, it first performs an automatic search over the memory base upon receiving a user query. This initial search is a coarse retrieval spanning all six memory components and returns high-level summaries rather than detailed content. The Chat Agent then analyzes the query to determine which memory components warrant more targeted searches and selects appropriate retrieval methods accordingly. After obtaining the relevant results, it consolidates the information and synthesizes the final response. Moreover, if the user's query involves updating memory-for example, providing new facts or corrections-the Chat Agent can interact directly with the corresponding Memory Managers to apply precise updates to specific memory components.
Conversational Retrieval Workflow
In this work, we introduce MIRIX, a novel memory architecture designed to enhance the longterm reasoning and personalization capabilities of LLM-based agents. Unlike existing memory systems that primarily rely on flat storage or limited memory types, MIRIX leverages a structured and compositional approach, incorporating six specialized memory components-Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault-managed by dedicated Memory Managers under the coordination of a Meta Memory Manager. To rigorously evaluate our system, we introduce a challenging multimodal benchmark based on high-resolution screenshots of real user activity, demonstrating that MIRIX achieves substantial gains in accuracy and storage efficiency compared to both retrieval-augmented generation and long-context baselines. Experiments on the LOCOMO benchmark confirm that MIRIX delivers state-of-the-art performance in long-form conversational settings. Finally, to make these capabilities accessible to a broader audience, we build and release a personal assistant application powered by MIRIX, allowing users to experience consistent, memoryenhanced interactions in everyday scenarios. We hope this work paves the way for more robust, scalable, and human-like memory systems for LLM-based agents. In the future, we aim to build more challenging real-world benchmarks to comprehensively evaluate our system and constantly improve MIRIX and the associated personal assistant application to deliver better experiences to the users.
Experiments
Experimental Setup
Datasets
ScreenshotVQA We collect a new dataset that contains three PhD students' activities. We created a script that takes a screenshot every second. If the image taken this second is too similar to the last one (similarity > 0.99) then we skip the current image. With this script, three PhD students majored in Computer Science and Physics run the program from a week to a month. The first student is a heavy computer user, which leads to 5,886 images in one day (03/09/2025). The second student uses the computer slightly lighter, leading to 18,178 images for 20 days (05/16/2025-06/06/2025). The third student is the lightest computer user, having 5349 images for over a month (05/02/2025-06/14/2025). Then, with these screenshots, we ask each student to manually create some questions and then we double-check the questions to make sure they are answerable. Eventually, we have 11 questions from the first user, 21 questions from the second user, and 55 questions from the third user.
LOCOMO Following Mem0 [5], we choose LOCOMO dataset for a vertical comparison between MIRIX and existing memory systems. LOCOMO has 10 conversations, where each conversation has 600 dialogues and 26000 tokens on average. There are averagely 200 questions for each conversation, which is suitable for the evaluation for memory systems where we are reuiqred to inject the conversation into the memory and then answer the 200 questions according to the obtained memory. The questions can be classified into multiple categories: single-hop, multi-hop, temporal and open-domain. The dataset also has another category called "adversarial", which is used to test
whether the system can identify unanswerable questions. Following Mem0 [5], we exclude this category in our setting to provide fair comparisons with the earlier methods.
ScreenshotVQA
For ScreenshotVQA , since existing memory systems such as Letta, Mem0 still lack the ability to process multimodal input, we omit the comparisons with them. We consider the following baselines:
Gemini . As a long-context baseline, Gemini directly ingests the full set of screenshots to answer the questions. Because the original high-resolution screenshots exceed the model's context window, we resize them to 256×256 pixels, enabling approximately 3,600 images to fit into a single prompt. We then include the most recent 3,600 images in the prompt and query the model for an answer. Specifically, we use the gemini-2.5-flash-preview-04-17 model.
SigLIP . SigLIP serves as a retrieval-augmented baseline. We first use SigLIP to identify the top 50 most relevant images for each query and then prompt Gemini to generate the answer based on these retrieved images. In this setup, we employ google/siglip-so400m-patch14-384 as the retriever and gemini-2.5-flash-preview-04-17 as the language model.
We evaluate all methods using two metrics: (1) Accuracy , measured via an LLM-as-a-Judge approach comparing the generated responses to ground-truth answers and reporting the average accuracy across the three participants; and (2) Storage , defined as follows-For Gemini, we report the total size of the images after resizing to 256×256. For SigLIP, we report the total size of the original retrieved images in their native 2K-4K resolution. For MIRIX, we use sqlite as the storage backend and report the size of the sqlite.db file containing all extracted information.
The results are summarized in Table 1. As shown, MIRIX significantly outperforms all existing baselines while requiring substantially less storage. Specifically, because SigLIP retains the full set of high-resolution images, its storage footprint is very large, corresponding to the total size of all images in their original 2K-4K resolution. The long-context baseline (Gemini) reduces this by resizing images to 256×256 pixels, but still incurs considerable storage overhead. In contrast, MIRIX avoids storing raw images altogether and instead maintains only the compact sqlite database of
1 https://github.com/mem0ai/mem0/tree/main/evaluation
Table 1: Experimental Results on ScreenshotVQA.
Table 2: LLM-as-a-Judge scores (%, higher is better) for each question type in the LOCOMO dataset. As mentioned in Mem0 [5], as the average length in this dataset is only 9k, Full-Context is essentially the upper-bound. Thus recovering the performance of Full-Context shows the advancements of MIRIX.
extracted information, resulting in a much smaller storage size. Compared to retrieval-augmented generation (RAG) baselines, MIRIX achieves a 35% improvement in accuracy while reducing storage requirements by 99.9%. Relative to the long-context Gemini baseline, MIRIX yields a 410% improvement in accuracy with a 93.3% reduction in storage.
LOCOMO
Evaluation Metrics and Implementation Details
Evaluation Metrics In this paper, for both of the above datasets, we mainly consider the metric LLM-as-a-Judge . Specifically, we use GPT-4.1 as the judge to look at the question, answer, and response to predict whether the response addresses the question successfully.
Implementation Details for ScreenshotVQA Across all experiments, we use the gemini-2.5-flash-preview-04-17 model as the backbone. We selected Gemini because it integrates seamlessly with Google Cloud, enabling asynchronous image uploads and retrieval. This significantly accelerates processing, as each step requires multiple function calls: one to the meta memory manager and between zero and six to the other memory managers.
Implementation Details for LOCOMO For MIRIX, since every agent needs to call many functions to successfully insert the information into the memory systems, it requires the language model to have strong abilities on function calling. To this end, we use gpt-4.1-mini as our backbone model as it shows a stronger ability than gpt-4o-mini in terms of function calling. This is also revealled in Berkeley Function Calling Benchmark [36] where gpt-4o-mini has multi-turn overall acc as 22.12, vs the acc of gpt-4.1-mini being 29.75. For LOCOMO dataset, we run the baselines LangMem, RAG-500, Mem0 using the code provided in Mem0 1 with setting MODEL in the .env file as gpt-4.1-mini . Then for Zep, we use the code from their official repo 2 and replace the backbone model with gpt-4.1-mini . For all the baselines, we run their code once, while for MIRIX and Full-Context method, we run three times to report the average scores. We provide the results of each single run in Appendix A. The complete code for our evaluation and the predicted results from various baselines and MIRIX are provided in the public_evaluation branch 3 in our official repository.
Evaluation Metrics
Evaluation Metrics In this paper, for both of the above datasets, we mainly consider the metric LLM-as-a-Judge . Specifically, we use GPT-4.1 as the judge to look at the question, answer, and response to predict whether the response addresses the question successfully.
Implementation Details for ScreenshotVQA Across all experiments, we use the gemini-2.5-flash-preview-04-17 model as the backbone. We selected Gemini because it integrates seamlessly with Google Cloud, enabling asynchronous image uploads and retrieval. This significantly accelerates processing, as each step requires multiple function calls: one to the meta memory manager and between zero and six to the other memory managers.
Implementation Details for LOCOMO For MIRIX, since every agent needs to call many functions to successfully insert the information into the memory systems, it requires the language model to have strong abilities on function calling. To this end, we use gpt-4.1-mini as our backbone model as it shows a stronger ability than gpt-4o-mini in terms of function calling. This is also revealled in Berkeley Function Calling Benchmark [36] where gpt-4o-mini has multi-turn overall acc as 22.12, vs the acc of gpt-4.1-mini being 29.75. For LOCOMO dataset, we run the baselines LangMem, RAG-500, Mem0 using the code provided in Mem0 1 with setting MODEL in the .env file as gpt-4.1-mini . Then for Zep, we use the code from their official repo 2 and replace the backbone model with gpt-4.1-mini . For all the baselines, we run their code once, while for MIRIX and Full-Context method, we run three times to report the average scores. We provide the results of each single run in Appendix A. The complete code for our evaluation and the predicted results from various baselines and MIRIX are provided in the public_evaluation branch 3 in our official repository.
Implementation Details for ScreenshotVQA
For ScreenshotVQA , since existing memory systems such as Letta, Mem0 still lack the ability to process multimodal input, we omit the comparisons with them. We consider the following baselines:
Gemini . As a long-context baseline, Gemini directly ingests the full set of screenshots to answer the questions. Because the original high-resolution screenshots exceed the model's context window, we resize them to 256×256 pixels, enabling approximately 3,600 images to fit into a single prompt. We then include the most recent 3,600 images in the prompt and query the model for an answer. Specifically, we use the gemini-2.5-flash-preview-04-17 model.
SigLIP . SigLIP serves as a retrieval-augmented baseline. We first use SigLIP to identify the top 50 most relevant images for each query and then prompt Gemini to generate the answer based on these retrieved images. In this setup, we employ google/siglip-so400m-patch14-384 as the retriever and gemini-2.5-flash-preview-04-17 as the language model.
We evaluate all methods using two metrics: (1) Accuracy , measured via an LLM-as-a-Judge approach comparing the generated responses to ground-truth answers and reporting the average accuracy across the three participants; and (2) Storage , defined as follows-For Gemini, we report the total size of the images after resizing to 256×256. For SigLIP, we report the total size of the original retrieved images in their native 2K-4K resolution. For MIRIX, we use sqlite as the storage backend and report the size of the sqlite.db file containing all extracted information.
The results are summarized in Table 1. As shown, MIRIX significantly outperforms all existing baselines while requiring substantially less storage. Specifically, because SigLIP retains the full set of high-resolution images, its storage footprint is very large, corresponding to the total size of all images in their original 2K-4K resolution. The long-context baseline (Gemini) reduces this by resizing images to 256×256 pixels, but still incurs considerable storage overhead. In contrast, MIRIX avoids storing raw images altogether and instead maintains only the compact sqlite database of
1 https://github.com/mem0ai/mem0/tree/main/evaluation
Table 1: Experimental Results on ScreenshotVQA.
Table 2: LLM-as-a-Judge scores (%, higher is better) for each question type in the LOCOMO dataset. As mentioned in Mem0 [5], as the average length in this dataset is only 9k, Full-Context is essentially the upper-bound. Thus recovering the performance of Full-Context shows the advancements of MIRIX.
extracted information, resulting in a much smaller storage size. Compared to retrieval-augmented generation (RAG) baselines, MIRIX achieves a 35% improvement in accuracy while reducing storage requirements by 99.9%. Relative to the long-context Gemini baseline, MIRIX yields a 410% improvement in accuracy with a 93.3% reduction in storage.
Implementation Details for LOCOMO
Evaluation Metrics In this paper, for both of the above datasets, we mainly consider the metric LLM-as-a-Judge . Specifically, we use GPT-4.1 as the judge to look at the question, answer, and response to predict whether the response addresses the question successfully.
Implementation Details for ScreenshotVQA Across all experiments, we use the gemini-2.5-flash-preview-04-17 model as the backbone. We selected Gemini because it integrates seamlessly with Google Cloud, enabling asynchronous image uploads and retrieval. This significantly accelerates processing, as each step requires multiple function calls: one to the meta memory manager and between zero and six to the other memory managers.
Implementation Details for LOCOMO For MIRIX, since every agent needs to call many functions to successfully insert the information into the memory systems, it requires the language model to have strong abilities on function calling. To this end, we use gpt-4.1-mini as our backbone model as it shows a stronger ability than gpt-4o-mini in terms of function calling. This is also revealled in Berkeley Function Calling Benchmark [36] where gpt-4o-mini has multi-turn overall acc as 22.12, vs the acc of gpt-4.1-mini being 29.75. For LOCOMO dataset, we run the baselines LangMem, RAG-500, Mem0 using the code provided in Mem0 1 with setting MODEL in the .env file as gpt-4.1-mini . Then for Zep, we use the code from their official repo 2 and replace the backbone model with gpt-4.1-mini . For all the baselines, we run their code once, while for MIRIX and Full-Context method, we run three times to report the average scores. We provide the results of each single run in Appendix A. The complete code for our evaluation and the predicted results from various baselines and MIRIX are provided in the public_evaluation branch 3 in our official repository.
Experimental Results on ScreenshotVQA
For ScreenshotVQA , since existing memory systems such as Letta, Mem0 still lack the ability to process multimodal input, we omit the comparisons with them. We consider the following baselines:
Gemini . As a long-context baseline, Gemini directly ingests the full set of screenshots to answer the questions. Because the original high-resolution screenshots exceed the model's context window, we resize them to 256×256 pixels, enabling approximately 3,600 images to fit into a single prompt. We then include the most recent 3,600 images in the prompt and query the model for an answer. Specifically, we use the gemini-2.5-flash-preview-04-17 model.
SigLIP . SigLIP serves as a retrieval-augmented baseline. We first use SigLIP to identify the top 50 most relevant images for each query and then prompt Gemini to generate the answer based on these retrieved images. In this setup, we employ google/siglip-so400m-patch14-384 as the retriever and gemini-2.5-flash-preview-04-17 as the language model.
We evaluate all methods using two metrics: (1) Accuracy , measured via an LLM-as-a-Judge approach comparing the generated responses to ground-truth answers and reporting the average accuracy across the three participants; and (2) Storage , defined as follows-For Gemini, we report the total size of the images after resizing to 256×256. For SigLIP, we report the total size of the original retrieved images in their native 2K-4K resolution. For MIRIX, we use sqlite as the storage backend and report the size of the sqlite.db file containing all extracted information.
The results are summarized in Table 1. As shown, MIRIX significantly outperforms all existing baselines while requiring substantially less storage. Specifically, because SigLIP retains the full set of high-resolution images, its storage footprint is very large, corresponding to the total size of all images in their original 2K-4K resolution. The long-context baseline (Gemini) reduces this by resizing images to 256×256 pixels, but still incurs considerable storage overhead. In contrast, MIRIX avoids storing raw images altogether and instead maintains only the compact sqlite database of
1 https://github.com/mem0ai/mem0/tree/main/evaluation
Table 1: Experimental Results on ScreenshotVQA.
Table 2: LLM-as-a-Judge scores (%, higher is better) for each question type in the LOCOMO dataset. As mentioned in Mem0 [5], as the average length in this dataset is only 9k, Full-Context is essentially the upper-bound. Thus recovering the performance of Full-Context shows the advancements of MIRIX.
extracted information, resulting in a much smaller storage size. Compared to retrieval-augmented generation (RAG) baselines, MIRIX achieves a 35% improvement in accuracy while reducing storage requirements by 99.9%. Relative to the long-context Gemini baseline, MIRIX yields a 410% improvement in accuracy with a 93.3% reduction in storage.
Experimental Results on LOCOMO
We compare MIRIX against the following baselines:
A-Mem [35]: A memory system that builds Zettelkasten-style knowledge graphs from user-agent interactions, dynamically linking notes using embedding similarity and LLM reasoning.
LangMem 4 : LangChain's long-term memory module that extracts and stores salient facts from conversations for later retrieval through retrievers like FAISS or Chroma.
Zep [24]: A commercial memory API that constructs a temporal knowledge graph (Graphiti) over user conversations and metadata, designed for fast semantic querying.
Mem0 [5]: An open-source memory system that incrementally compresses and stores memory facts using LLM-based summarization, with an optional graph memory extension.
Memobase 5 : A profile-based memory module that tracks persistent user attributes and preferences to enable long-term personalization.
All baselines are re-implemented using the same backbone model ( gpt-4.1-mini ). We also report the results shown in Mem0 [5] where the backbone model is gpt-4o-mini . The results are shown in Table 2. From the table, we observe the following:
4 https://github.com/langchain-ai/langmem
Overall: MIRIX achieves the highest average J score, outperforming all baselines by a significant margin. It improves upon the strongest open-source competitor, LangMem, by over 8 points.
Single-Hop and Temporal: On fact lookup and temporal-ordering tasks, MIRIX shows significantly better performances than baselines, which validates the effectiveness of our hierarchical memory storage. We note that there is a minor gap between MIRIX and the Full-Context baseline on SingleHop questions. Upon reviewing the results, we identified a major reason for this. In some cases, the questions are ambiguous. For example, consider the question 'When is Melanie planning on going camping?' In the conversation history, Melanie stated 'We're thinking about going camping next month' in May, which would suggest June as the answer. However, she later mentioned in October, 'Absolutely! It really helps me reset and recharge. I love camping trips with my fam, 'cause nature brings such peace and serenity, ' referring to a more recent trip. Because MIRIX saves the consolidated event 'On 19 October 2023, Melanie and her family went camping after their road trip, ' it tends to prioritize the confirmed occurrence over the earlier plan, leading to discrepancies when the question expects the planned date rather than the actual event. Other similar ambiguities contribute to the slightly lower performance of MIRIX on Single-Hop questions.
Multi-Hop: MIRIX demonstrates the largest gains in this category, outperforming all baselines by over 24 points. For example, in questions such as 'Where did Caroline move from 4 years ago?' , where the correct answer is Sweden, the supporting evidence is dispersed across multiple parts of the conversation. One part may state 'Caroline moved from her hometown 4 years ago' , while an earlier statement establishes that 'Caroline's hometown is Sweden' . For multi-hop questions like these, MIRIX achieves better performance because it explicitly stores the consolidated event 'Caroline moved from her hometown, Sweden, 4 years ago' , removing the need to stitch together partial information at query time. In contrast, full-context methods must first retrieve the partial answer 'hometown' and then figure out separately that 'hometown' refers to 'Swedon' . These multi-hop questions might be easy for reasoning models like OpenAI-O3, but for non-reasoning models such as gpt-4.1-mini , this additional reasoning step might fail, leading to slightly inferior performance compared with MIRIX.
Open-Domain: While MIRIX performs well, the margin between ours and the baselines is narrower. This category of questions usually asks the agent 'what if' questions, requiring the agent to infer across longer terms. The gap between MIRIX and Full-Context method show the inherent limitation of RAG methods, which is the lack of global understanding. While MIRIX is no longer simple RAG, we still rely on RAG to retrieve important information in the memory, which might lead to the bottleneck of our agent in this category.
In summary, these results demonstrate that MIRIX delivers state-of-the-art performance on LOCOMO while remaining highly efficient and modular. Its component-specific memory management and intelligent routing are particularly effective for long-range multi-hop reasoning.
Related Work
Memory-Augmented Large Language Models Agrowing body of work focuses on building latentspace memory systems, as characterized in M+[31], where transformer architectures are modified to support memory augmentation. These memory components can reside in various latent forms, including model parameters[30], external memory matrices [8, 21], hidden states [2, 10, 28, 31], soft prompts [3], and key-value caches [38, 9, 16, 37]. While these approaches demonstrate promising results and continue to advance the frontier of latent-space memory, most require retraining the model [28, 31, 21], making them incompatible with powerful closed-source models like GPT-4 or DeepSeek-R1. Additionally, methods based on key-value caching rely on preserving past keys and values, functioning more as long-context methods rather than true memory systems that support abstraction, consolidation, and reasoning over stored experiences [29].
Memory-Augmented LLM Agents Token-level memory remains the dominant approach in current LLM agents [31], where past conversational content is stored in raw text form within external databases. Notable examples include commercial systems such as Zep [24], Mem0 [5], and MemGPT [22]. These agents perform well on long-term conversational benchmarks [19, 34] and document-based retrieval tasks. However, they often fall short in real-world applications due to
simplistic memory architectures. Most notably, the absence of modular memory components hinders effective memory routing and leads to inefficiencies in retrieval and usage.
Various Memory Types Cognitive science broadly categorizes memory into short-term (working) memory and long-term memory [26]. In the context of LLMs, short-term memory is often mapped to the input context window [14], while long-term memory becomes a catch-all category for any information outside the context [31]. To address this limitation, recent works have proposed finergrained memory architectures. For example, Pink et al. [23] emphasizes the importance of episodic memory in LLM agents. Other systems incorporate both episodic and semantic memory [1, 15, 27]. Semantic memory has also been highlighted as critical for real-world reasoning and abstraction [33, 13]. In addition, procedural memory, responsible for learned skills and routine tasks, has been identified as another crucial component [33]. Despite these advancements, existing methods stop at identifying individual memory types, and they are not formed into a comprehensive memory system.
Multi-Agent Systems Rather than relying on a monolithic agent, recent advances explore multiagent frameworks where specialized agents coordinate to accomplish complex tasks. Early systems like AutoGPT and BabyAGI [6, 7] adopt an autonomous planning-execution loop while maintaining a shared memory log. More recent designs introduce role specialization: MetaGPT [11] mimics a software development team structure, and AgentVerse [4] assigns agents to specific roles such as planning or evaluation. Cognitive theories also support modularity in memory, particularly distinctions between episodic and semantic types, as emphasized in Liao et al. [17]. MIRIX builds on these ideas by deploying eight specialized agents, each managing a distinct memory type (e.g., episodic, semantic, procedural), and coordinating to process multi-modal inputs effectively.
Memory-Augmented Large Language Models
Memory-Augmented LLM Agents
We design a modular memory architecture consisting of six distinct components: Core Memory , Episodic Memory , Semantic Memory , Procedural Memory , Resource Memory , and the Knowledge Vault . Each component is structurally and functionally tailored to capture different aspects of user interaction and world knowledge, enabling the agent to retrieve, reason, and act effectively across time and tasks.
Core Memory Core Memory stores high-priority, persistent information that should always remain visible to the agent when engaging with the user. Inspired by the design of MemGPT [22], this memory is divided into two primary blocks: persona and human . The persona block encodes the identity, tone, or behavior profile of the agent, while the human block stores enduring facts about the user such as their name, preferences, and self-identifying attributes (e.g., 'User's name is David', 'User enjoys Japanese cuisine'). When the memory size exceeds 90% of capacity, the system triggers a controlled rewrite process to maintain compactness without losing critical information.
Episodic Memory Episodic Memory captures time-stamped events and temporally grounded interactions that reflect the user's behavior, experiences, or activities. It functions as a structured log or calendar, enabling the agent to reason about user routines, recency, and context-aware followups [17, 23, 18]. Each entry is defined by the following fields: event_type (e.g., user_message , inferred_result , system_notification ), summary (a concise natural language description of the event), details (extended contextual information, including dialog excerpts or inferred states), actor (the origin of the event, either user or assistant ), and timestamp (e.g., '2025-03-05 10:15'). This structure allows the agent to temporally index memory and track change over time, such as identifying ongoing tasks or following up on pending actions.
Figure 5: Demonstration of Active Retrieval.
Semantic Memory Semantic Memory maintains abstract knowledge and factual information that is independent of specific times or events. This component serves as a knowledge base for general concepts, entities, and relationships - whether about the world or the user's social graph. For example, it may store entries like 'Harry Potter is written by J.K. Rowling' or 'John is a friend of the user who enjoys jogging and lives in San Francisco.' Each entry includes a name (the concept or entity identifier), summary (a concise definition or relationship statement), details (expanded background or contextual explanation), and source (e.g., user_provided , Wikipedia , or inferred from conversation). Unlike episodic memory, semantic entries are intended to persist unless conceptually overwritten and support reasoning over social, geographic, or commonsense knowledge.
Procedural Memory Procedural Memory stores structured, goal-directed processes such as how-to guides, operational workflows, and interactive scripts. These are neither time-sensitive (episodic) nor abstract facts (semantic) but instead represent actionable knowledge that can be invoked to assist the user with complex tasks. Typical examples include 'how to file a travel reimbursement form,' 'steps to set up a Zoom meeting,' or 'how to book a restaurant via OpenTable.' Each entry includes an entry_type ( workflow , guide , or script ), a description of the goal or function, and steps expressed as a list of instructions (optionally in JSON or structured format). This memory component supports instructional planning, automation, and decomposition of user goals into sub-tasks.
Resource Memory Resource Memory handles full or partial documents, transcripts, or multi-modal files that the user is actively engaged with but do not fit into other memory categories. For instance, if the user is reading a friend's detailed picnic plan or a project proposal document, the agent can store and retrieve that information from Resource Memory. This component enables context continuity in long-running tasks. Each entry includes a title (resource name), summary (brief overview and context), resource_type (e.g., doc , markdown , pdf_text , image , voice_transcript ), and the full or excerpted content . This design enables the agent to reference previously seen material, quote from documents, or search within them to aid the user's workflow.
Knowledge Vault The Knowledge Vault serves as a secure repository for verbatim and sensitive information such as credentials, addresses, contact information, and API keys. These entries are not typically relevant to conversation-level reasoning but are crucial for performing authenticated tasks or storing long-term identifiers. Each entry includes an entry_type (e.g., credential , bookmark , contact_info , api_key ), a source (e.g., user_provided , github ), a sensitivity level ( low , medium , high ), and the actual secret_value . Entries with high sensitivity are protected via access control and excluded from casual retrieval to prevent misuse or leakage.
Various Memory Types
We envision a future where personal memory-collected and structured through AI agents-becomes a new digital asset class. In the AI era, memory is no longer just a passive log of past events, but an active, evolving knowledge base that can be shared, personalized, and monetized. The Agent Memory Marketplace is our proposal for a decentralized ecosystem where memory is exchanged, reused, and built collaboratively through AI agents.
We begin with a core belief: human memory will become the most valuable and irreplaceable asset in the age of AI . Unlike static data, memory encompasses lived experiences, subjective context,
preferences, and interactions-making it deeply personal, yet highly reusable by intelligent systems. the marketplace is structured into three key layers:
- AI Agents Infrastructure . Our technology provides infrastructure for lifetime intelligent, interactive agents. Examples include (1) Personal AI Assistants & Companions: Tailored agents that continuously learn and evolve with the user. (2) AI Wearables: AI agents embedded in decentralized physical infrastructure (e.g., smart glasses, AI pins) that extend memory capture to the real world. (3) Multi-Agent Systems: Collaborative agents with shared memory access, enabling coordination and collective intelligence.
- Privacy-Preserving Memory Infrastructure To support trust and adoption, we aim to build in a robust privacy layer: (1) Encryption Layer: All memories are stored using end-to-end encryption. (2) Privacy Control: Fine-grained permissions allow users to choose which parts of their memory to share, trade, or restrict. (3) Decentralized Storage: Memories are stored in a distributed, censorship-resistant infrastructure.
- Memory Marketplace and Social Function : A peer-to-peer ecosystem for sharing, aggregating, and trading memories encoded in AI agents. Use cases include: (1) Memory Social/Trading: Tokenized exchange of memories (e.g., productivity hacks, niche workflows, or life advice). (2) Expert Communities: Collective memory-building for domain-specific expertise such as finance, education, or pet care. (3) Fan Economy and Dating Applications: Users can subscribe to AI personas based on influencer or celebrity memories, allowing fans to interact with memory-rich digital replicas of their favorite personalities. This technology also enables the creation of AI clones for accelerated dating and matching processes, where users can engage with AI agents in interactions before meeting in person.
In summary, we envision a future where personal memory transcends its traditional role to become an active, valuable digital asset. By combining lifelong AI agents, privacy-preserving infrastructure, and a decentralized marketplace, we aim to create an ecosystem where memories can be securely captured, meaningfully shared, and collectively advanced. This approach not only empowers individuals to benefit from their own experiences but also unlocks new possibilities for collaboration, personalization, and economic value in the AI era.
Multi-Agent Systems
To manage the dynamic and heterogeneous nature of user interactions, we adopt a modular multiagent architecture. This system orchestrates input processing, memory updating, and information retrieval across six distinct memory components through a coordinated and efficient workflow. The overall system is governed by a central Meta Memory Manager and a set of specialized Memory Managers , each responsible for maintaining one memory type.
Memory Update Workflow As illustrated in Figure 7, when new input is received from the user, the system first automatically performs a search over the memory base. The retrieved information, together with the user input, is passed to the Meta Memory Manager . The Meta Memory Manager then analyzes the content and determines which memory components are relevant, routing the input to the corresponding Memory Managers . These Memory Managers update their respective memories in parallel while ensuring that redundant information is avoided within each memory type. After
Figure 7: The Workflow of Responding to the User's Query.
completing the updates, they report back to the Meta Memory Manager , which finally sends an acknowledgment confirming that the memory update process is complete.
Conversational Retrieval Workflow For interactive dialogues, the Chat Agent manages natural language communication with the user. To ground its responses in prior knowledge, it first performs an automatic search over the memory base upon receiving a user query. This initial search is a coarse retrieval spanning all six memory components and returns high-level summaries rather than detailed content. The Chat Agent then analyzes the query to determine which memory components warrant more targeted searches and selects appropriate retrieval methods accordingly. After obtaining the relevant results, it consolidates the information and synthesizes the final response. Moreover, if the user's query involves updating memory-for example, providing new facts or corrections-the Chat Agent can interact directly with the corresponding Memory Managers to apply precise updates to specific memory components.
Conclusion and Future Work
In this work, we introduce MIRIX, a novel memory architecture designed to enhance the longterm reasoning and personalization capabilities of LLM-based agents. Unlike existing memory systems that primarily rely on flat storage or limited memory types, MIRIX leverages a structured and compositional approach, incorporating six specialized memory components-Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault-managed by dedicated Memory Managers under the coordination of a Meta Memory Manager. To rigorously evaluate our system, we introduce a challenging multimodal benchmark based on high-resolution screenshots of real user activity, demonstrating that MIRIX achieves substantial gains in accuracy and storage efficiency compared to both retrieval-augmented generation and long-context baselines. Experiments on the LOCOMO benchmark confirm that MIRIX delivers state-of-the-art performance in long-form conversational settings. Finally, to make these capabilities accessible to a broader audience, we build and release a personal assistant application powered by MIRIX, allowing users to experience consistent, memoryenhanced interactions in everyday scenarios. We hope this work paves the way for more robust, scalable, and human-like memory systems for LLM-based agents. In the future, we aim to build more challenging real-world benchmarks to comprehensively evaluate our system and constantly improve MIRIX and the associated personal assistant application to deliver better experiences to the users.
Experiments
Full Experimental Results with Different Runs
Although memory capabilities of AI agents are gaining increasing attention, existing solutions remain fundamentally limited. Most rely on flat, narrowly scoped memory components, constraining their ability to personalize, abstract, and reliably recall user-specific information over time. To this end, we introduce MIRIX, a modular, multi-agent memory system that redefines the future of AI memory by solving the field’s most critical challenge: enabling language models to truly remember. Unlike prior approaches, MIRIX transcends text to embrace rich visual and multimodal experiences, making memory genuinely useful in real-world scenarios. MIRIX consists of six distinct, carefully structured memory types: Core, Episodic, Semantic, Procedural, Resource Memory, and Knowledge Vault, coupled with a multi-agent framework that dynamically controls and coordinates updates and retrieval. This design enables agents to persist, reason over, and accurately retrieve diverse, long-term user data at scale. We validate MIRIX in two demanding settings. First, on ScreenshotVQA, a challenging multimodal benchmark comprising nearly 20,000 high-resolution computer screenshots per sequence, requiring deep contextual understanding and where no existing memory systems can be applied, MIRIX achieves 35% higher accuracy than the RAG baseline while reducing storage requirements by 99.9%. Second, on LOCOMO, a long-form conversation benchmark with single-modal textual input, MIRIX attains state-of-the-art performance of 85.4%, far surpassing existing baselines. These results show that MIRIX sets a new performance standard for memory-augmented LLM Agents. To allow users to experience our memory system, we provide a packaged application powered by MIRIX. It monitors the screen in real time, builds a personalized memory base, and offers intuitive visualization and secure local storage to ensure privacy.
Recent advancements in large language model (LLM) agents have focused primarily on improving their capabilities in complex task execution—ranging from code debugging and repository management to autonomous web browsing. While these functionalities are crucial, another foundational yet underexplored dimension is memory: the ability of agents to persist, retrieve, and utilize past user-specific information over time. Human cognition relies heavily on memory—recalling conversations, recognizing patterns, and adapting behavior based on prior experience. Analogously, memory mechanisms in LLM agents are essential for delivering consistent, personalized interactions, learning from feedback, and avoiding repetitive queries. However, most LLM-based personal assistants remain stateless beyond their current prompt window, retaining no lasting memory unless context is explicitly re-provided. This limitation hinders their long-term usability, especially in real-world settings where users expect assistants to evolve, recall, and personalize over time Wang et al. (2024b).
To address this, a range of memory-augmented systems have been proposed. One common approach is the use of knowledge graphs, as seen in systems like Zep (Rasmussen et al., 2025) and Cognee (Markovic et al., 2025). These frameworks are well-suited for representing structured relationships between entities but struggle to model sequential events, emotional states, full-length documents, or multi-modal inputs such as images. Another approach involves flattened memory architectures that store and retrieve textual chunks using vector databases. Examples include Letta (Packer et al., 2023), Mem0 (Chhikara et al., 2025), and ChatGPT’s memory system. Letta divides its memory into components—recall memory for conversation history, core memory for preferences, and archival memory for long documents—while ChatGPT focuses primarily on core and recall memories. Mem0 adopts a memory system that contains flattened facts distilled from the user inputs, which serves similar roles as Letta’s archival memory while being more distilled. While prevalent, these memory systems face several challenges: (1) Lack of compositional memory structure: Most approaches store all historical data in a single flat store without routing into specialized memory types (e.g., procedural, episodic, semantic), making retrieval inefficient and less accurate. (2) Poor multi-modal support: Text-centric memory mechanisms fail when the majority of the input is non-verbal (e.g., images, interface layouts, maps). (3) Scalability and abstraction: Storing raw inputs, especially images, leads to prohibitive memory requirements, with no effective abstraction layer to summarize and retain only salient information.
To address the limitations of existing memory systems, we argue that effective Routing and Retrieving are the key capabilities a memory-augmented agent must possess. Most current systems focus primarily on Short-Term and Long-Term Memory (Packer et al., 2023; Westhäußer et al., 2025; Shan et al., 2025), with some incorporating Mid-Term Memory (Kang et al., 2025). In contrast, we draw inspiration from works that explore more specialized memory types, including episodic memory (Pink et al., 2025; Liu et al., 2025; Anokhin et al., 2024), semantic memory (Li and Li, 2024b), and procedural memory (Wheeler and Jeunen, 2025). Building on these foundations, we propose a more comprehensive architecture consisting of six memory components: Core Memory, Episodic Memory, Semantic Memory, Procedural Memory, Resource Memory, and the Knowledge Vault (As shown in Figure 1). Episodic Memory stores user-specific events and experiences; Semantic Memory captures concepts and named entities (e.g., the meaning of a new phrase or an understanding of a person); Procedural Memory records step-by-step instructions for performing tasks. Resource Memory is designed to store documents, files, and other media shared by the user. Knowledge Vault holds critical verbatim information that must be preserved exactly, such as addresses, phone numbers, email accounts, and other sensitive facts.
Each memory component is internally organized using a hierarchical structure. For example, Episodic Memory includes fields such as summary and details, while Semantic Memory organizes information by name and description. Managing this structured and heterogeneous memory is challenging for a single agent. Therefore, we adopt a multi-agent architecture: six Memory Managers for six memory components, and a Meta Memory Manager responsible for task routing. While this memory system can be plugged and connected with other existing agents, we build an extra Chat Agent to demonstrate how we can interact with an agent that has access to our memories. This memory system, which we call MIRIX, is modular and designed as a comprehensive and full memory system for LLM-based agents. Moreover, when interacting with the Chat Agent, we propose an Active Retrieval mechanism, where the agent is required to generate a topic before answering the question or executing the next step, and the retrieved information is inputted into the model in the system prompt. Meanwhile, we design multiple retrieval tools so that the agent can choose appropriate ones in response to different situations.
We evaluate MIRIX in two experimental settings. First, we introduce a challenging benchmark that requires extracting information and building memory from multimodal input. To this end, we collect between 5,000 and 20,000 high-resolution screenshots spanning one month of computer usage from three PhD students and construct evaluation questions grounded in their visual activity history. This setting demands sophisticated memory modeling and exceeds the capacity of existing long-context models. For example, each screenshot ranges from 2K to 4K resolution depending on the user’s monitor, which limits Gemini to processing no more than 500 images at full resolution or approximately 3,600 images when resized to 256×256 pixels. Second, following Mem0 (Chhikara et al., 2025), we assess performance on the LOCOMO dataset (Maharana et al., 2024), which comprises long-form, multi-turn conversations. While each conversation is relatively short (around 26,000 tokens on average), we constrain the Chat Agent to answer questions using only retrieved memories, without access to the original conversation transcripts. This setup enables us to evaluate whether our system can effectively distill and route essential information into memory. In terms of results, because no existing memory systems can handle such a large volume of multimodal input, we compare against retrieval-augmented generation (RAG) baselines and long-context baselines. MIRIX achieves a 35% improvement over RAG baselines while reducing storage requirements by 99.9%, and a 410% improvement over long-context baselines with a 93.3% reduction in storage. On LOCOMO, MIRIX reaches state-of-the-art performance with an overall accuracy of 85.38%, outperforming the best existing method by 8.0% and approaching the upper bound set by long-context models. To make our memory system more accessible, we also developed a personal assistant powered by MIRIX. With the user’s permission to capture screen content, the assistant continuously builds a memory from screenshots and can answer any questions related to this accumulated information. This represents our effort to bring advanced memory capabilities to everyone, enabling users to experience the benefits of our system firsthand.
Our contributions are summarized as follows:
We analyze the limitations of existing memory architectures for LLM-based agents and propose a novel memory system composed of six specialized components and eight different agents.
We introduce a new benchmark in which the agent must interpret a large collection of screenshots to build an understanding of user behavior—posing a significant challenge to the memory.
Through experiments on our new benchmark and existing benchmark LOCOMO, we demonstrate that MIRIX significantly outperform existing memory systems.
To make our approach accessible, we build and release a personal assistant application powered by MIRIX, enabling users to experience advanced memory capabilities in real-world scenarios.
To demonstrate the full functionality of our memory system, we developed a cross-platform application using React-Electron for the frontend and Uvicorn as the backend server. Moreover, we release the obtained file for direct installation. Within the application, users can activate screen monitoring by selecting the ScreenShots tab. This enables the agent to observe the user’s screen activity and dynamically update its memory, gradually building a contextual understanding of the user over time. With the obtained memory, the agent can answer questions related to the memories.
The application captures a screenshot every 1.5 seconds. To reduce redundancy, images that are visually similar to previously captured ones are discarded. Once 20 unique screenshots are collected, the memory update process is triggered—typically around every 60 seconds. During this process, relevant information is extracted and incorporated into the system’s memory components. To reduce latency in processing user screenshots, we adopt a streaming upload strategy. Instead of batching and sending 20 images at once, we upload each screenshot immediately upon receiving it from the front-end. By leveraging the Gemini API—which supports loading images via Google Cloud URLs—we can efficiently transmit visual data without waiting for the full batch to accumulate. This approach significantly reduces end-to-end latency from approximately 50 seconds (as observed when using GPT-4 with direct image upload) to under 5 seconds using Gemini.
The chat interface allows users to interact with the agent, which has full access to its accumulated memories. As illustrated in Figure 2, users can query the assistant about past activities, and the agent can respond based on its memory content, enabling more informed and personalized interactions.
After observing the screen for a sufficient period, the agent organizes its knowledge into structured memory components. An example of Semantic Memory (organized into a tree structure) is shown in Figure 3. We also provide a list view of the memories. An example of Procedural Memory is shown in Figure 4.
The wearable device market has seen rapid growth in recent years, driven by the increasing demand for intelligent, always-available personal assistants. Products like AI-powered glasses (e.g., Meta Ray-Ban, XREAL Air) and AI pins (e.g., Humane, Rabbit R1) aim to integrate seamless interaction into daily life through voice commands, visual capture, and real-time feedback. However, these devices often lack a long-term memory component that allows them to evolve with the user—retaining useful information over time, adapting to personal routines, and referencing past interactions in context-aware ways.
Our memory system is well-suited for integration into such wearable devices. By continuously collecting and processing data streams such as audio, visual scenes, and user queries, our system enables real-time memory formation. For instance, AI glasses equipped with our memory systems can automatically summarize meetings, remember frequently visited places, recognize recurring visual patterns, and recall previous conversations or tasks. With MIRIX, it can also evolve with the user and build a memory that is specifically made for the user. Moreover, the system’s modular memory architecture—including procedural, episodic, semantic, and resource memory—aligns naturally with the needs of lightweight, on-the-go devices. Procedural memory enables the assistant to learn user habits (e.g., daily routes, meeting structures), while semantic memory stores general knowledge about the user’s preferences, environment, and routines. Episodic memory captures time-stamped, situational experiences, and can be queried for recalling specific events (e.g., “What did I see at the conference last week?”). Semantic Memory, on the other hand, can help organize the clients someone has seen over the past week and list them in a tree structure with details of their discussions.
Given the constraints of wearable hardware (limited compute and storage), our design also supports hybrid on-device/cloud memory management. Critical information in Knowledge Vault can be stored locally while large-scale memories such as Resource Memory can be offloaded and retrieved from the cloud on demand.
In summary, our memory system serves as a cognitive backbone for wearable AI agents—enabling personalization, continuity, and intelligence at the edge. As the wearable market matures, embedding persistent, structured memory will be a key differentiator for next-generation AI assistants.
We envision a future where personal memory—collected and structured through AI agents—becomes a new digital asset class. In the AI era, memory is no longer just a passive log of past events, but an active, evolving knowledge base that can be shared, personalized, and monetized. The Agent Memory Marketplace is our proposal for a decentralized ecosystem where memory is exchanged, reused, and built collaboratively through AI agents.
We begin with a core belief: human memory will become the most valuable and irreplaceable asset in the age of AI. Unlike static data, memory encompasses lived experiences, subjective context, preferences, and interactions—making it deeply personal, yet highly reusable by intelligent systems. the marketplace is structured into three key layers:
AI Agents Infrastructure. Our technology provides infrastructure for lifetime intelligent, interactive agents. Examples include (1) Personal AI Assistants & Companions: Tailored agents that continuously learn and evolve with the user. (2) AI Wearables: AI agents embedded in decentralized physical infrastructure (e.g., smart glasses, AI pins) that extend memory capture to the real world. (3) Multi-Agent Systems: Collaborative agents with shared memory access, enabling coordination and collective intelligence.
Privacy-Preserving Memory Infrastructure To support trust and adoption, we aim to build in a robust privacy layer: (1) Encryption Layer: All memories are stored using end-to-end encryption. (2) Privacy Control: Fine-grained permissions allow users to choose which parts of their memory to share, trade, or restrict. (3) Decentralized Storage: Memories are stored in a distributed, censorship-resistant infrastructure.
Memory Marketplace and Social Function: A peer-to-peer ecosystem for sharing, aggregating, and trading memories encoded in AI agents. Use cases include: (1) Memory Social/Trading: Tokenized exchange of memories (e.g., productivity hacks, niche workflows, or life advice). (2) Expert Communities: Collective memory-building for domain-specific expertise such as finance, education, or pet care. (3) Fan Economy and Dating Applications: Users can subscribe to AI personas based on influencer or celebrity memories, allowing fans to interact with memory-rich digital replicas of their favorite personalities. This technology also enables the creation of AI clones for accelerated dating and matching processes, where users can engage with AI agents in interactions before meeting in person.
In summary, we envision a future where personal memory transcends its traditional role to become an active, valuable digital asset. By combining lifelong AI agents, privacy-preserving infrastructure, and a decentralized marketplace, we aim to create an ecosystem where memories can be securely captured, meaningfully shared, and collectively advanced. This approach not only empowers individuals to benefit from their own experiences but also unlocks new possibilities for collaboration, personalization, and economic value in the AI era.
We design a modular memory architecture consisting of six distinct components: Core Memory, Episodic Memory, Semantic Memory, Procedural Memory, Resource Memory, and the Knowledge Vault. Each component is structurally and functionally tailored to capture different aspects of user interaction and world knowledge, enabling the agent to retrieve, reason, and act effectively across time and tasks.
Core Memory stores high-priority, persistent information that should always remain visible to the agent when engaging with the user. Inspired by the design of MemGPT (Packer et al., 2023), this memory is divided into two primary blocks: persona and human. The persona block encodes the identity, tone, or behavior profile of the agent, while the human block stores enduring facts about the user such as their name, preferences, and self-identifying attributes (e.g., “User’s name is David”, “User enjoys Japanese cuisine”). When the memory size exceeds 90% of capacity, the system triggers a controlled rewrite process to maintain compactness without losing critical information.
Episodic Memory captures time-stamped events and temporally grounded interactions that reflect the user’s behavior, experiences, or activities. It functions as a structured log or calendar, enabling the agent to reason about user routines, recency, and context-aware follow-ups (Liao et al., 2024; Pink et al., 2025; Liu et al., 2025). Each entry is defined by the following fields: event_type (e.g., user_message, inferred_result, system_notification), summary (a concise natural language description of the event), details (extended contextual information, including dialog excerpts or inferred states), actor (the origin of the event, either user or assistant), and timestamp (e.g., “2025-03-05 10:15”). This structure allows the agent to temporally index memory and track change over time, such as identifying ongoing tasks or following up on pending actions.
Semantic Memory maintains abstract knowledge and factual information that is independent of specific times or events. This component serves as a knowledge base for general concepts, entities, and relationships — whether about the world or the user’s social graph. For example, it may store entries like “Harry Potter is written by J.K. Rowling” or “John is a friend of the user who enjoys jogging and lives in San Francisco.” Each entry includes a name (the concept or entity identifier), summary (a concise definition or relationship statement), details (expanded background or contextual explanation), and source (e.g., user_provided, Wikipedia, or inferred from conversation). Unlike episodic memory, semantic entries are intended to persist unless conceptually overwritten and support reasoning over social, geographic, or commonsense knowledge.
Procedural Memory stores structured, goal-directed processes such as how-to guides, operational workflows, and interactive scripts. These are neither time-sensitive (episodic) nor abstract facts (semantic) but instead represent actionable knowledge that can be invoked to assist the user with complex tasks. Typical examples include “how to file a travel reimbursement form,” “steps to set up a Zoom meeting,” or “how to book a restaurant via OpenTable.” Each entry includes an entry_type (workflow, guide, or script), a description of the goal or function, and steps expressed as a list of instructions (optionally in JSON or structured format). This memory component supports instructional planning, automation, and decomposition of user goals into sub-tasks.
Resource Memory handles full or partial documents, transcripts, or multi-modal files that the user is actively engaged with but do not fit into other memory categories. For instance, if the user is reading a friend’s detailed picnic plan or a project proposal document, the agent can store and retrieve that information from Resource Memory. This component enables context continuity in long-running tasks. Each entry includes a title (resource name), summary (brief overview and context), resource_type (e.g., doc, markdown, pdf_text, image, voice_transcript), and the full or excerpted content. This design enables the agent to reference previously seen material, quote from documents, or search within them to aid the user’s workflow.
The Knowledge Vault serves as a secure repository for verbatim and sensitive information such as credentials, addresses, contact information, and API keys. These entries are not typically relevant to conversation-level reasoning but are crucial for performing authenticated tasks or storing long-term identifiers. Each entry includes an entry_type (e.g., credential, bookmark, contact_info, api_key), a source (e.g., user_provided, github), a sensitivity level (low, medium, high), and the actual secret_value. Entries with high sensitivity are protected via access control and excluded from casual retrieval to prevent misuse or leakage.
In many memory-augmented systems (e.g., Mem0 (Chhikara et al., 2025), MemGPT (Packer et al., 2023)), memory retrieval must be explicitly triggered. Otherwise, the language model often defaults to its parametric knowledge, which may be outdated or incorrect. For example, suppose the user previously said, “The CEO of Twitter is Linda Yaccarino,” and this information was saved in memory. A few days later, when this message is no longer present in the conversation history, the user might ask, “Who is the CEO of Twitter?” In this case, the language model may rely on outdated knowledge and incorrectly respond with “Elon Musk.” While explicitly instructing the model to “search your memory” can mitigate such errors, doing so repeatedly is impractical in natural conversations.
To address this, we propose an Active Retrieval mechanism. As illustrated in Figure 5, the system operates in two stages: first, the agent generates a current topic based on the input context; second, this topic is used to retrieve relevant memories from each memory component. Retrieved results are then injected into the system prompt. For instance, given the query “Who is the CEO of Twitter?”, the agent may infer the topic “CEO of Twitter”, which is then used to retrieve the top-10 most relevant entries from each of the six memory components. Retrieved content is tagged according to its source, such as <episodic_memory>… </episodic_memory>, ensuring the model is aware of both the content and its origin. This automatic retrieval pipeline eliminates the need for explicit user prompts to trigger memory access and ensures the model can incorporate up-to-date, personalized, or contextual information during response generation.
In addition to Active Retrieval, we support multiple retrieval functions, including embedding_match, bm25_match, and string_match. We are actively expanding this set with more diverse and specialized retrieval strategies, ensuring that each method is well-differentiated so the agent can choose the most appropriate one to invoke based on context.
To manage the dynamic and heterogeneous nature of user interactions, we adopt a modular multi-agent architecture. This system orchestrates input processing, memory updating, and information retrieval across six distinct memory components through a coordinated and efficient workflow. The overall system is governed by a central Meta Memory Manager and a set of specialized Memory Managers, each responsible for maintaining one memory type.
As illustrated in Figure 7, when new input is received from the user, the system first automatically performs a search over the memory base. The retrieved information, together with the user input, is passed to the Meta Memory Manager. The Meta Memory Manager then analyzes the content and determines which memory components are relevant, routing the input to the corresponding Memory Managers. These Memory Managers update their respective memories in parallel while ensuring that redundant information is avoided within each memory type. After completing the updates, they report back to the Meta Memory Manager, which finally sends an acknowledgment confirming that the memory update process is complete.
For interactive dialogues, the Chat Agent manages natural language communication with the user. To ground its responses in prior knowledge, it first performs an automatic search over the memory base upon receiving a user query. This initial search is a coarse retrieval spanning all six memory components and returns high-level summaries rather than detailed content. The Chat Agent then analyzes the query to determine which memory components warrant more targeted searches and selects appropriate retrieval methods accordingly. After obtaining the relevant results, it consolidates the information and synthesizes the final response. Moreover, if the user’s query involves updating memory—for example, providing new facts or corrections-the Chat Agent can interact directly with the corresponding Memory Managers to apply precise updates to specific memory components.
We collect a new dataset that contains three PhD students’ activities. We created a script that takes a screenshot every second. If the image taken this second is too similar to the last one (similarity >> 0.99) then we skip the current image. With this script, three PhD students majored in Computer Science and Physics run the program from a week to a month. The first student is a heavy computer user, which leads to 5,886 images in one day (03/09/2025). The second student uses the computer slightly lighter, leading to 18,178 images for 20 days (05/16/2025-06/06/2025). The third student is the lightest computer user, having 5349 images for over a month (05/02/2025-06/14/2025). Then, with these screenshots, we ask each student to manually create some questions and then we double-check the questions to make sure they are answerable. Eventually, we have 11 questions from the first user, 21 questions from the second user, and 55 questions from the third user.
Following Mem0 (Chhikara et al., 2025), we choose LOCOMO dataset for a vertical comparison between MIRIX and existing memory systems. LOCOMO has 10 conversations, where each conversation has 600 dialogues and 26000 tokens on average. There are averagely 200 questions for each conversation, which is suitable for the evaluation for memory systems where we are reuiqred to inject the conversation into the memory and then answer the 200 questions according to the obtained memory. The questions can be classified into multiple categories: single-hop, multi-hop, temporal and open-domain. The dataset also has another category called "adversarial", which is used to test whether the system can identify unanswerable questions. Following Mem0 (Chhikara et al., 2025), we exclude this category in our setting to provide fair comparisons with the earlier methods.
In this paper, for both of the above datasets, we mainly consider the metric LLM-as-a-Judge. Specifically, we use GPT-4.1 as the judge to look at the question, answer, and response to predict whether the response addresses the question successfully.
Across all experiments, we use the gemini-2.5-flash-preview-04-17 model as the backbone. We selected Gemini because it integrates seamlessly with Google Cloud, enabling asynchronous image uploads and retrieval. This significantly accelerates processing, as each step requires multiple function calls: one to the meta memory manager and between zero and six to the other memory managers.
For MIRIX, since every agent needs to call many functions to successfully insert the information into the memory systems, it requires the language model to have strong abilities on function calling. To this end, we use gpt-4.1-mini as our backbone model as it shows a stronger ability than gpt-4o-mini in terms of function calling. This is also revealled in Berkeley Function Calling Benchmark (Yan et al., 2024) where gpt-4o-mini has multi-turn overall acc as 22.12, vs the acc of gpt-4.1-mini being 29.75. For LOCOMO dataset, we run the baselines LangMem, RAG-500, Mem0 using the code provided in Mem0 111https://github.com/mem0ai/mem0/tree/main/evaluation with setting MODEL in the .env file as gpt-4.1-mini. Then for Zep, we use the code from their official repo222https://github.com/getzep/zep-papers?ref=www.getzep.com/blog and replace the backbone model with gpt-4.1-mini. For all the baselines, we run their code once, while for MIRIX and Full-Context method, we run three times to report the average scores. We provide the results of each single run in Appendix A. The complete code for our evaluation and the predicted results from various baselines and MIRIX are provided in the public_evaluation branch333https://github.com/Mirix-AI/MIRIX/tree/public_evaluation in our official repository.
For ScreenshotVQA, since existing memory systems such as Letta, Mem0 still lack the ability to process multimodal input, we omit the comparisons with them. We consider the following baselines:
Gemini. As a long-context baseline, Gemini directly ingests the full set of screenshots to answer the questions. Because the original high-resolution screenshots exceed the model’s context window, we resize them to 256×256 pixels, enabling approximately 3,600 images to fit into a single prompt. We then include the most recent 3,600 images in the prompt and query the model for an answer. Specifically, we use the gemini-2.5-flash-preview-04-17 model.
SigLIP. SigLIP serves as a retrieval-augmented baseline. We first use SigLIP to identify the top 50 most relevant images for each query and then prompt Gemini to generate the answer based on these retrieved images. In this setup, we employ google/siglip-so400m-patch14-384 as the retriever and gemini-2.5-flash-preview-04-17 as the language model.
We evaluate all methods using two metrics: (1) Accuracy, measured via an LLM-as-a-Judge approach comparing the generated responses to ground-truth answers and reporting the average accuracy across the three participants; and (2) Storage, defined as follows—For Gemini, we report the total size of the images after resizing to 256×256. For SigLIP, we report the total size of the original retrieved images in their native 2K–4K resolution. For MIRIX, we use sqlite as the storage backend and report the size of the sqlite.db file containing all extracted information.
The results are summarized in Table 1. As shown, MIRIX significantly outperforms all existing baselines while requiring substantially less storage. Specifically, because SigLIP retains the full set of high-resolution images, its storage footprint is very large, corresponding to the total size of all images in their original 2K–4K resolution. The long-context baseline (Gemini) reduces this by resizing images to 256×256 pixels, but still incurs considerable storage overhead. In contrast, MIRIX avoids storing raw images altogether and instead maintains only the compact sqlite database of extracted information, resulting in a much smaller storage size. Compared to retrieval-augmented generation (RAG) baselines, MIRIX achieves a 35% improvement in accuracy while reducing storage requirements by 99.9%. Relative to the long-context Gemini baseline, MIRIX yields a 410% improvement in accuracy with a 93.3% reduction in storage.
A-Mem (Xu et al., 2025): A memory system that builds Zettelkasten-style knowledge graphs from user-agent interactions, dynamically linking notes using embedding similarity and LLM reasoning.
LangMem444https://github.com/langchain-ai/langmem: LangChain’s long-term memory module that extracts and stores salient facts from conversations for later retrieval through retrievers like FAISS or Chroma.
Memobase555https://github.com/memodb-io/memobase: A profile-based memory module that tracks persistent user attributes and preferences to enable long-term personalization.
All baselines are re-implemented using the same backbone model (gpt-4.1-mini). We also report the results shown in Mem0 (Chhikara et al., 2025) where the backbone model is gpt-4o-mini. The results are shown in Table 2. From the table, we observe the following:
Overall: MIRIX achieves the highest average J score, outperforming all baselines by a significant margin. It improves upon the strongest open-source competitor, LangMem, by over 8 points.
Single-Hop and Temporal: On fact lookup and temporal-ordering tasks, MIRIX shows significantly better performances than baselines, which validates the effectiveness of our hierarchical memory storage. We note that there is a minor gap between MIRIX and the Full-Context baseline on Single-Hop questions. Upon reviewing the results, we identified a major reason for this. In some cases, the questions are ambiguous. For example, consider the question “When is Melanie planning on going camping?” In the conversation history, Melanie stated “We’re thinking about going camping next month” in May, which would suggest June as the answer. However, she later mentioned in October, “Absolutely! It really helps me reset and recharge. I love camping trips with my fam, ’cause nature brings such peace and serenity,” referring to a more recent trip. Because MIRIX saves the consolidated event “On 19 October 2023, Melanie and her family went camping after their road trip,” it tends to prioritize the confirmed occurrence over the earlier plan, leading to discrepancies when the question expects the planned date rather than the actual event. Other similar ambiguities contribute to the slightly lower performance of MIRIX on Single-Hop questions.
Multi-Hop: MIRIX demonstrates the largest gains in this category, outperforming all baselines by over 24 points. For example, in questions such as “Where did Caroline move from 4 years ago?”, where the correct answer is Sweden, the supporting evidence is dispersed across multiple parts of the conversation. One part may state “Caroline moved from her hometown 4 years ago”, while an earlier statement establishes that “Caroline’s hometown is Sweden”. For multi-hop questions like these, MIRIX achieves better performance because it explicitly stores the consolidated event “Caroline moved from her hometown, Sweden, 4 years ago”, removing the need to stitch together partial information at query time. In contrast, full-context methods must first retrieve the partial answer “hometown” and then figure out separately that “hometown” refers to “Swedon”. These multi-hop questions might be easy for reasoning models like OpenAI-O3, but for non-reasoning models such as gpt-4.1-mini, this additional reasoning step might fail, leading to slightly inferior performance compared with MIRIX.
Open-Domain: While MIRIX performs well, the margin between ours and the baselines is narrower. This category of questions usually asks the agent “what if” questions, requiring the agent to infer across longer terms. The gap between MIRIX and Full-Context method show the inherent limitation of RAG methods, which is the lack of global understanding. While MIRIX is no longer simple RAG, we still rely on RAG to retrieve important information in the memory, which might lead to the bottleneck of our agent in this category.
In summary, these results demonstrate that MIRIX delivers state-of-the-art performance on LOCOMO while remaining highly efficient and modular. Its component-specific memory management and intelligent routing are particularly effective for long-range multi-hop reasoning.
A growing body of work focuses on building latent-space memory systems, as characterized in M+(Wang et al., 2025), where transformer architectures are modified to support memory augmentation. These memory components can reside in various latent forms, including model parameters(Wang et al., 2024c), external memory matrices (Das et al., 2024; Munkhdalai et al., 2024), hidden states (Bulatov et al., 2022; He et al., 2024; Wang et al., 2024a, 2025), soft prompts (Burtsev and Sapunov, 2020), and key-value caches (Zhong et al., 2023; Eyuboglu et al., 2025; Li et al., 2024; Zhang et al., 2023). While these approaches demonstrate promising results and continue to advance the frontier of latent-space memory, most require retraining the model (Wang et al., 2024a, 2025; Munkhdalai et al., 2024), making them incompatible with powerful closed-source models like GPT-4 or DeepSeek-R1. Additionally, methods based on key-value caching rely on preserving past keys and values, functioning more as long-context methods rather than true memory systems that support abstraction, consolidation, and reasoning over stored experiences (Wang et al., 2024b).
Token-level memory remains the dominant approach in current LLM agents (Wang et al., 2025), where past conversational content is stored in raw text form within external databases. Notable examples include commercial systems such as Zep (Rasmussen et al., 2025), Mem0 (Chhikara et al., 2025), and MemGPT (Packer et al., 2023). These agents perform well on long-term conversational benchmarks (Maharana et al., 2024; Wu et al., 2024) and document-based retrieval tasks. However, they often fall short in real-world applications due to simplistic memory architectures. Most notably, the absence of modular memory components hinders effective memory routing and leads to inefficiencies in retrieval and usage.
Cognitive science broadly categorizes memory into short-term (working) memory and long-term memory (Sridhar et al., 2023). In the context of LLMs, short-term memory is often mapped to the input context window (Li and Li, 2024a), while long-term memory becomes a catch-all category for any information outside the context (Wang et al., 2025). To address this limitation, recent works have proposed finer-grained memory architectures. For example, Pink et al. (2025) emphasizes the importance of episodic memory in LLM agents. Other systems incorporate both episodic and semantic memory (Anokhin et al., 2024; Li and Li, 2024b; Tulving, 1985). Semantic memory has also been highlighted as critical for real-world reasoning and abstraction (Wheeler and Jeunen, 2025; Kim et al., 2023). In addition, procedural memory, responsible for learned skills and routine tasks, has been identified as another crucial component (Wheeler and Jeunen, 2025). Despite these advancements, existing methods stop at identifying individual memory types, and they are not formed into a comprehensive memory system.
Rather than relying on a monolithic agent, recent advances explore multi-agent frameworks where specialized agents coordinate to accomplish complex tasks. Early systems like AutoGPT and BabyAGI (Community, 2023a, b) adopt an autonomous planning-execution loop while maintaining a shared memory log. More recent designs introduce role specialization: MetaGPT (Hong et al., 2023) mimics a software development team structure, and AgentVerse (Chen et al., 2024) assigns agents to specific roles such as planning or evaluation. Cognitive theories also support modularity in memory, particularly distinctions between episodic and semantic types, as emphasized in Liao et al. (2024). MIRIX builds on these ideas by deploying eight specialized agents, each managing a distinct memory type (e.g., episodic, semantic, procedural), and coordinating to process multi-modal inputs effectively.
In this work, we introduce MIRIX, a novel memory architecture designed to enhance the long-term reasoning and personalization capabilities of LLM-based agents. Unlike existing memory systems that primarily rely on flat storage or limited memory types, MIRIX leverages a structured and compositional approach, incorporating six specialized memory components—Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault—managed by dedicated Memory Managers under the coordination of a Meta Memory Manager. To rigorously evaluate our system, we introduce a challenging multimodal benchmark based on high-resolution screenshots of real user activity, demonstrating that MIRIX achieves substantial gains in accuracy and storage efficiency compared to both retrieval-augmented generation and long-context baselines. Experiments on the LOCOMO benchmark confirm that MIRIX delivers state-of-the-art performance in long-form conversational settings. Finally, to make these capabilities accessible to a broader audience, we build and release a personal assistant application powered by MIRIX, allowing users to experience consistent, memory-enhanced interactions in everyday scenarios. We hope this work paves the way for more robust, scalable, and human-like memory systems for LLM-based agents. In the future, we aim to build more challenging real-world benchmarks to comprehensively evaluate our system and constantly improve MIRIX and the associated personal assistant application to deliver better experiences to the users.
We run MIRIX and Full-Context with gpt-4.1-mini three times and we report the full results in Table 3. There are variations across different runs, while MIRIX consistently achieves state-of-the-art results. Full predicted results and LLM-Judge scores are provided in the folder https://github.com/ Mirix-AI/MIRIX/tree/public_evaluation/public_evaluations/evaluation_metrics .
Table: S4.T1: Experimental Results on ScreenshotVQA.
| Acc ↑\uparrow | Storage↓\downarrow | Acc ↑\uparrow | Storage↓\downarrow | Acc ↑\uparrow | Storage↓\downarrow | Acc ↑\uparrow | Storage↓\downarrow | |
|---|---|---|---|---|---|---|---|---|
| Gemini | 0.0000 | 142.10MB | 0.0952 | 438.86MB | 0.2545 | 129.14MB | 0.1166 | 236.70MB |
| SigLIP@50 | 0.3636 | 22.55GB | 0.4138 | 19.88GB | 0.5455 | 2.82GB | 0.4410 | 15.07GB |
| MIRIX | 0.5455 | 20.57MB | 0.5667 | 19.83MB | 0.6727 | 7.28MB | 0.5950 | 15.89MB |
Table: S4.T2: LLM-as-a-Judge scores (%, higher is better) for each question type in the LOCOMO dataset. As mentioned in Mem0 (Chhikara et al., 2025), as the average length in this dataset is only 9k, Full-Context is essentially the upper-bound. Thus recovering the performance of Full-Context shows the advancements of MIRIX.
| Method | Single Hop | Multi-Hop | Open Domain | Temporal | Overall | |
| gpt-4o-mini | A-Mem | 39.79 | 18.85 | 54.05 | 49.91 | 48.38 |
| LangMem | 62.23 | 47.92 | 71.12 | 23.43 | 58.10 | |
| OpenAI | 63.79 | 42.92 | 62.29 | 21.71 | 52.90 | |
| Mem0 | 67.13 | 51.15 | 72.93 | 55.51 | 66.88 | |
| Mem0g | 65.71 | 47.19 | 75.71 | 58.13 | 68.44 | |
| Memobase | 63.83 | 52.08 | 71.82 | 80.37 | 70.91 | |
| Zep | 74.11 | 66.04 | 67.71 | 79.76 | 75.14 | |
| gpt-4.1-mini | LangMem | 74.47 | 61.06 | 67.71 | 86.92 | 78.05 |
| RAG-500 | 37.94 | 37.69 | 48.96 | 61.83 | 51.62 | |
| Zep | 79.43 | 69.16 | 73.96 | 83.33 | 79.09 | |
| Mem0 | 62.41 | 57.32 | 44.79 | 66.47 | 62.47 | |
| MIRIX | 85.11 | 83.70 | 65.62 | 88.39 | 85.38 | |
| Full-Context | 88.53 | 77.70 | 71.88 | 92.70 | 87.52 |
The six memory components of MIRIX, each providing specialized functionality.
Chat Window
Tree Structure of an Example Semantic Memory. In this example, the user’s Semantic Memory, which stores new concepts and their relationships, is organized hierarchically into multiple categories, such as Social Network and Favorites. Within Favorites, the memory is further divided into more specific classes, including Sports, Pets and Music.
List View of an Example Procedural Memory.
Demonstration of Active Retrieval.
The Workflow of Memory Update.
Table 3: LLM-as-a-Judge scores (%, higher is better) for each question type in the LOCOMO dataset. As mentioned in Mem0 [5], as the average length in this dataset is only 9k, Full-Context is almost the upper-bound. Thus recovering the performance of Full-Context shows the advancements of MIRIX. For Zep, we only achieve 49.09 overall score with gpt-4.1-mini using the implementation from mem0, we are afraid there might be errors in their implementation, so we skip the results here.
| Student 1 | Student 1 | Student 2 | Student 2 | Student 3 | Student 3 | Overall | Overall | |
|---|---|---|---|---|---|---|---|---|
| Acc ↑ | Storage ↓ | Acc ↑ | Storage ↓ | Acc ↑ | Storage ↓ | Acc ↑ | Storage ↓ | |
| Gemini | 0.0000 | 142.10MB | 0.0952 | 438.86MB | 0.2545 | 129.14MB | 0.1166 | 236.70MB |
| SigLIP@50 | 0.3636 | 22.55GB | 0.4138 | 19.88GB | 0.5455 | 2.82GB | 0.4410 | 15.07GB |
| MIRIX | 0.5455 | 20.57MB | 0.5667 | 19.83MB | 0.6727 | 7.28MB | 0.5950 | 15.89MB |
| Method | Single Hop | Multi-Hop | Open Domain | Temporal | Overall | |
|---|---|---|---|---|---|---|
| A-Mem | 39.79 | 18.85 | 54.05 | 49.91 | 48.38 | |
| LangMem | 62.23 | 47.92 | 71.12 | 23.43 | 58.1 | |
| OpenAI | 63.79 | 42.92 | 62.29 | 21.71 | 52.9 | |
| gpt-4o-mini | Mem0 | 67.13 | 51.15 | 72.93 | 55.51 | 66.88 |
| Mem0 g | 65.71 | 47.19 | 75.71 | 58.13 | 68.44 | |
| Memobase | 63.83 | 52.08 | 71.82 | 80.37 | 70.91 | |
| Zep | 74.11 | 66.04 | 67.71 | 79.76 | 75.14 | |
| LangMem | 74.47 | 61.06 | 67.71 | 86.92 | 78.05 | |
| RAG-500 | 37.94 | 37.69 | 48.96 | 61.83 | 51.62 | |
| Zep | 79.43 | 69.16 | 73.96 | 83.33 | 79.09 | |
| gpt-4.1-mini | Mem0 | 62.41 | 57.32 | 44.79 | 66.47 | 62.47 |
| MIRIX | 85.11 | 83.7 | 65.62 | 88.39 | 85.38 | |
| Full-Context | 88.53 | 77.7 | 71.88 | 92.7 | 87.52 |
| Method | Single Hop | Multi-Hop | Open Domain | Temporal | Overall |
|---|---|---|---|---|---|
| A-Mem | 39.79 | 18.85 | 54.05 | 49.91 | 48.38 |
| LangMem | 62.23 | 47.92 | 71.12 | 23.43 | 58.1 |
| OpenAI | 63.79 | 42.92 | 62.29 | 21.71 | 52.9 |
| Mem0 | 67.13 | 51.15 | 72.93 | 55.51 | 66.88 |
| Mem0 g | 65.71 | 47.19 | 75.71 | 58.13 | 68.44 |
| Memobase | 63.83 | 52.08 | 71.82 | 80.37 | 70.91 |
| Zep | 74.11 | 66.04 | 67.71 | 79.76 | 75.14 |
| Full-Context | 80.14 | 46.89 | 58.33 | 90.49 | 77.51 |
| LangMem | 76.24 | 63.24 | 63.54 | 88.11 | 79.22 |
| RAG-500 | 37.94 | 37.69 | 48.96 | 61.83 | 51.62 |
| Mem0 | 62.41 | 57.32 | 44.79 | 66.47 | 62.47 |
| MIRIX-Run1 | 85.46 | 80.06 | 64.58 | 87.16 | 83.98 |
| MIRIX-Run2 | 85.46 | 85.36 | 64.58 | 91.32 | 87.34 |
| MIRIX-Run3 | 84.4 | 85.67 | 67.71 | 86.68 | 84.82 |
| Full-Context-Run1 | 88.3 | 71.74 | 72.92 | 92.98 | 86.43 |
| Full-Context-Run2 | 88.29 | 81.31 | 71.88 | 92.5 | 88.13 |
| Full-Context-Run3 | 89.01 | 80.06 | 70.83 | 92.63 | 88 |
| Student 1 | Student 1 | Student 2 | Student 2 | Student 3 | Student 3 | Overall | Overall | |
|---|---|---|---|---|---|---|---|---|
| Acc ↑ | Storage ↓ | Acc ↑ | Storage ↓ | Acc ↑ | Storage ↓ | Acc ↑ | Storage ↓ | |
| Gemini | 0.0000 | 142.10MB | 0.0952 | 438.86MB | 0.2545 | 129.14MB | 0.1166 | 236.70MB |
| SigLIP@50 | 0.3636 | 22.55GB | 0.4138 | 19.88GB | 0.5455 | 2.82GB | 0.4410 | 15.07GB |
| MIRIX | 0.5455 | 20.57MB | 0.5667 | 19.83MB | 0.6727 | 7.28MB | 0.5950 | 15.89MB |
| Method | Single Hop | Multi-Hop | Open Domain | Temporal | Overall | |
|---|---|---|---|---|---|---|
| A-Mem | 39.79 | 18.85 | 54.05 | 49.91 | 48.38 | |
| LangMem | 62.23 | 47.92 | 71.12 | 23.43 | 58.1 | |
| OpenAI | 63.79 | 42.92 | 62.29 | 21.71 | 52.9 | |
| gpt-4o-mini | Mem0 | 67.13 | 51.15 | 72.93 | 55.51 | 66.88 |
| Mem0 g | 65.71 | 47.19 | 75.71 | 58.13 | 68.44 | |
| Memobase | 63.83 | 52.08 | 71.82 | 80.37 | 70.91 | |
| Zep | 74.11 | 66.04 | 67.71 | 79.76 | 75.14 | |
| LangMem | 74.47 | 61.06 | 67.71 | 86.92 | 78.05 | |
| RAG-500 | 37.94 | 37.69 | 48.96 | 61.83 | 51.62 | |
| Zep | 79.43 | 69.16 | 73.96 | 83.33 | 79.09 | |
| gpt-4.1-mini | Mem0 | 62.41 | 57.32 | 44.79 | 66.47 | 62.47 |
| MIRIX | 85.11 | 83.7 | 65.62 | 88.39 | 85.38 | |
| Full-Context | 88.53 | 77.7 | 71.88 | 92.7 | 87.52 |
| Method | Single Hop | Multi-Hop | Open Domain | Temporal | Overall |
|---|---|---|---|---|---|
| A-Mem | 39.79 | 18.85 | 54.05 | 49.91 | 48.38 |
| LangMem | 62.23 | 47.92 | 71.12 | 23.43 | 58.1 |
| OpenAI | 63.79 | 42.92 | 62.29 | 21.71 | 52.9 |
| Mem0 | 67.13 | 51.15 | 72.93 | 55.51 | 66.88 |
| Mem0 g | 65.71 | 47.19 | 75.71 | 58.13 | 68.44 |
| Memobase | 63.83 | 52.08 | 71.82 | 80.37 | 70.91 |
| Zep | 74.11 | 66.04 | 67.71 | 79.76 | 75.14 |
| Full-Context | 80.14 | 46.89 | 58.33 | 90.49 | 77.51 |
| LangMem | 76.24 | 63.24 | 63.54 | 88.11 | 79.22 |
| RAG-500 | 37.94 | 37.69 | 48.96 | 61.83 | 51.62 |
| Mem0 | 62.41 | 57.32 | 44.79 | 66.47 | 62.47 |
| MIRIX-Run1 | 85.46 | 80.06 | 64.58 | 87.16 | 83.98 |
| MIRIX-Run2 | 85.46 | 85.36 | 64.58 | 91.32 | 87.34 |
| MIRIX-Run3 | 84.4 | 85.67 | 67.71 | 86.68 | 84.82 |
| Full-Context-Run1 | 88.3 | 71.74 | 72.92 | 92.98 | 86.43 |
| Full-Context-Run2 | 88.29 | 81.31 | 71.88 | 92.5 | 88.13 |
| Full-Context-Run3 | 89.01 | 80.06 | 70.83 | 92.63 | 88 |
References
[MemoryNetwork] Weston, Jason, Chopra, Sumit, Bordes, Antoine. (2014). Memory networks. arXiv preprint arXiv:1410.3916.
[kNNLM] Khandelwal, Urvashi, Levy, Omer, Jurafsky, Dan, Zettlemoyer, Luke, Lewis, Mike. (2019). Generalization through memorization: Nearest neighbor language models. arXiv preprint arXiv:1911.00172.
[MemoryBank] Zhong, Wanjun, Guo, Lianghong, Gao, Qiqi, Wang, Yanlin. (2023). MemoryBank: Enhancing Large Language Models with Long-Term Memory. arXiv preprint arXiv:2305.10250.
[LLMCWM] Daliang Li, Ankit Singh Rawat, Manzil Zaheer, Xin Wang, Michal Lukasik, Andreas Veit, Felix X. Yu, Sanjiv Kumar. (2022). Large Language Models with Controllable Working Memory. CoRR.
[unlimiformer] Bertsch, Amanda, Alon, Uri, Neubig, Graham, Gormley, Matthew R. (2023). Unlimiformer: Long-range transformers with unlimited length input. arXiv preprint arXiv:2305.01625.
[alibi] Press, Ofir, Smith, Noah A, Lewis, Mike. (2021). Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409.
[longnet] Jiayu, Ding, Shuming, Ma, Li, Dong, Xingxing, Zhang, Shaohan, Huang, Wenhui, Wang, Furu Wei†. (2023). LONGNET: Scaling Transformers to 1,000,000,000 Tokens. arXiv preprint arXiv:2307.02486.
[memoryTransformer] Mikhail S. Burtsev, Grigory V. Sapunov. (2020). Memory Transformer. CoRR.
[LongMEM] Wang, Weizhi, Dong, Li, Cheng, Hao, Liu, Xiaodong, Yan, Xifeng, Gao, Jianfeng, Wei, Furu. (2023). Augmenting Language Models with Long-Term Memory. arXiv preprint arXiv:2306.07174.
[Memory-Enhanced-Transformer] Gianluca Moro, Luca Ragazzi, Lorenzo Valgimigli, Giacomo Frisoni, Claudio Sartori, Gustavo Marfia. (2023). Efficient Memory-Enhanced Transformer for Long-Document Summarization in Low-Resource Regimes. Sensors.
[HEPOS] Luyang Huang, Shuyang Cao, Nikolaus Nova Parulian, Heng Ji, Lu Wang. (2021). Efficient Attentions for Long Document Summarization. {NAACL-HLT.
[FMTforIC] Tongwei Lu, Jiarong Wang, Fen Min. (2023). Full-Memory Transformer for Image Captioning. Symmetry.
[MeshedTforIC] Marcella Cornia, Matteo Stefanini, Lorenzo Baraldi, Rita Cucchiara. (2020). Meshed-Memory Transformer for Image Captioning. {CVPR.
[ModelingHumanMemory] Yizhuo Li, Cewu Lu. (2022). Modeling Human Memory in Multi-Object Tracking with Transformers. {ICASSP.
[Memformer] Qingyang Wu, Zhenzhong Lan, Kun Qian, Jing Gu, Alborz Geramifard, Zhou Yu. (2022). Memformer: {A. {AACL/IJCNLP.
[RMT] Aydar Bulatov, Yuri Kuratov, Mikhail S. Burtsev. (2022). Recurrent Memory Transformer. NeurIPS.
[TMR] Rui Liu, Barzan Mozafari. (2022). Transformer with Memory Replay. {AAAI.
[GlobalMemoryTransformer] Arij Al Adel. (2022). Global memory transformer for processing long documents. CoRR.
[LifeLongPretraining] Xisen Jin, Dejiao Zhang, Henghui Zhu, Wei Xiao, Shang{-. (2022). Lifelong Pretraining: Continually Adapting Language Models to Emerging Corpora. {NAACL-HLT.
[TemporalWiki] Joel Jang, Seonghyeon Ye, Changho Lee, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Minjoon Seo. (2022). TemporalWiki: {A. {EMNLP.
[LLMEditing] Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, Ningyu Zhang. (2023). Editing Large Language Models: Problems, Methods, and Opportunities. CoRR.
[TPatcher] Huang, Zeyu, Shen, Yikang, Zhang, Xiaofeng, Zhou, Jie, Rong, Wenge, Xiong, Zhang. (2023). Transformer-Patcher: One Mistake worth One Neuron. arXiv preprint arXiv:2301.09785.
[PositionalInterpolation] Chen, Shouyuan, Wong, Sherman, Chen, Liangjian, Tian, Yuandong. (2023). Extending Context Window of Large Language Models via Positional Interpolation. arXiv preprint arXiv:2306.15595.
[CKL] Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stanley Jungkyu Choi, Minjoon Seo. (2022). Towards Continual Knowledge Learning of Language Models. {ICLR.
[longformer] Iz Beltagy, Matthew E. Peters, Arman Cohan. (2020). Longformer: The Long-Document Transformer. CoRR.
[lextransformer] Yutao Sun, Li Dong, Barun Patra, Shuming Ma, Shaohan Huang, Alon Benhaim, Vishrav Chaudhary, Xia Song, Furu Wei. (2023). A Length-Extrapolatable Transformer. {ACL.
[SparserTransformer] Rewon Child, Scott Gray, Alec Radford, Ilya Sutskever. (2019). Generating Long Sequences with Sparse Transformers. CoRR.
[bigbird] Zaheer, Manzil, Guruganesh, Guru, Dubey, Kumar Avinava, Ainslie, Joshua, Alberti, Chris, Ontanon, Santiago, Pham, Philip, Ravula, Anirudh, Wang, Qifan, Yang, Li, others. (2020). Big bird: Transformers for longer sequences. Advances in neural information processing systems.
[reformer] Kitaev, Nikita, Kaiser, {\L. (2020). Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451.
[linformer] Wang, Sinong, Li, Belinda Z, Khabsa, Madian, Fang, Han, Ma, Hao. (2020). Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768.
[E2EMN] Sukhbaatar, Sainbayar, Weston, Jason, Fergus, Rob, others. (2015). End-to-end memory networks. Advances in neural information processing systems.
[MemoryTransformerWithHierarchicalAttention] Al Adel, Arij, Burtsev, Mikhail S. (2021). Memory transformer with hierarchical attention for long document processing. 2021 International Conference Engineering and Telecommunication (En&T).
[Llama2Long] Xiong, Wenhan, Liu, Jingyu, Molybog, Igor, Zhang, Hejia, Bhargava, Prajjwal, Hou, Rui, Martin, Louis, Rungta, Rashi, Sankararaman, Karthik Abinav, Oguz, Barlas, others. (2023). Effective long-context scaling of foundation models. arXiv preprint arXiv:2309.16039.
[longbench] Bai, Yushi, Lv, Xin, Zhang, Jiajie, Lyu, Hongchang, Tang, Jiankai, Huang, Zhidian, Du, Zhengxiao, Liu, Xiao, Zeng, Aohan, Hou, Lei, others. (2023). LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding. arXiv preprint arXiv:2308.14508.
[chapterbreak] Sun, Simeng, Thai, Katherine, Iyyer, Mohit. (2022). Chapterbreak: A challenge dataset for long-range language models. arXiv preprint arXiv:2204.10878.
[fot] Tworkowski, Szymon, Staniszewski, Konrad, Pacek, Miko{\l. (2023). Focused transformer: Contrastive training for context scaling. arXiv preprint arXiv:2307.03170.
[SERAC] Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D. Manning, Chelsea Finn. (2022). Memory-Based Model Editing at Scale. {ICML.
[memprompt] Madaan, Aman, Tandon, Niket, Clark, Peter, Yang, Yiming. (2022). Memory-assisted prompt editing to improve gpt-3 after deployment. arXiv preprint arXiv:2201.06009.
[ike] Zheng, Ce, Li, Lei, Dong, Qingxiu, Fan, Yuxuan, Wu, Zhiyong, Xu, Jingjing, Chang, Baobao. (2023). Can We Edit Factual Knowledge by In-Context Learning?. arXiv preprint arXiv:2305.12740.
[mello] Zhong, Zexuan, Wu, Zhengxuan, Manning, Christopher D, Potts, Christopher, Chen, Danqi. (2023). MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions. arXiv preprint arXiv:2305.14795.
[calinet] Qingxiu Dong, Damai Dai, Yifan Song, Jingjing Xu, Zhifang Sui, Lei Li. (2022). Calibrating Factual Knowledge in Pretrained Language Models. {EMNLP.
[grace] Hartvigsen, Thomas, Sankaranarayanan, Swami, Palangi, Hamid, Kim, Yoon, Ghassemi, Marzyeh. (2022). Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors. arXiv preprint arXiv:2211.11031.
[KN] Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, Furu Wei. (2022). Knowledge Neurons in Pretrained Transformers. {ACL.
[ROME] Meng, Kevin, Bau, David, Andonian, Alex, Belinkov, Yonatan. (2022). Locating and editing factual associations in GPT. Advances in Neural Information Processing Systems.
[memit] Kevin Meng, Arnab Sen Sharma, Alex J. Andonian, Yonatan Belinkov, David Bau. (2023). Mass-Editing Memory in a Transformer. {ICLR.
[pmet] Li, Xiaopeng, Li, Shasha, Song, Shezheng, Yang, Jing, Ma, Jun, Yu, Jie. (2023). PMET: Precise Model Editing in a Transformer. arXiv preprint arXiv:2308.08742.
[KE] Nicola De Cao, Wilker Aziz, Ivan Titov. (2021). Editing Factual Knowledge in Language Models. {EMNLP.
[mend] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, Christopher D. Manning. (2022). Fast Model Editing at Scale. {ICLR.
[llama2] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton{-. (2023). Llama 2: Open Foundation and Fine-Tuned Chat Models. CoRR.
[openllama] Geng, Xinyang, Liu, Hao. OpenLLaMA: An Open Reproduction of LLaMA.
[zsre] Omer Levy, Minjoon Seo, Eunsol Choi, Luke Zettlemoyer. (2017). Zero-Shot Relation Extraction via Reading Comprehension. Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), Vancouver, Canada, August 3-4, 2017. doi:10.18653/V1/K17-1034.
[ModifyingMemories] Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix X. Yu, Sanjiv Kumar. (2020). Modifying Memories in Transformer Models. CoRR.
[c4] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu. (2020). Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. J. Mach. Learn. Res..
[together2023redpajama] Together Computer. RedPajama: an Open Dataset for Training Large Language Models.
[openlm2023openllama] Geng, Xinyang, Liu, Hao. OpenLLaMA: An Open Reproduction of LLaMA.
[longlora] Chen, Yukang, Qian, Shengju, Tang, Haotian, Lai, Xin, Liu, Zhijian, Han, Song, Jia, Jiaya. (2023). Longlora: Efficient fine-tuning of long-context large language models. arXiv preprint arXiv:2309.12307.
[ret-llm] Modarressi, Ali, Imani, Ayyoob, Fayyaz, Mohsen, Sch{. (2023). RET-LLM: Towards a General Read-Write Memory for Large Language Models. arXiv preprint arXiv:2305.14322.
[kv-cache-survey] Jiayi Yuan, Hongyi Liu, Shaochen Zhong, Yu{-. (2024). {KV. {EMNLP.
[h2o] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher R{'{e. (2023). {H2O:. NeurIPS.
[lscs] Yu Wang, Chi Han, Tongtong Wu, Xiaoxin He, Wangchunshu Zhou, Nafis Sadeq, Xiusi Chen, Zexue He, Wei Wang, Gholamreza Haffari, Heng Ji, Julian J. McAuley. (2024). Towards LifeSpan Cognitive Systems. CoRR.
[memoryllm] Yu Wang, Yifan Gao, Xiusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, Jingbo Shang, Julian J. McAuley. (2024). {MEMORYLLM:. {ICML.
[fedorenko2024language] Fedorenko, Evelina, Piantadosi, Steven T, Gibson, Edward AF. (2024). Language is primarily a tool for communication rather than thought. Nature.
[zhou2023recurrentgpt] Zhou, Wangchunshu, Jiang, Yuchen Eleanor, Cui, Peng, Wang, Tiannan, Xiao, Zhenxin, Hou, Yifan, Cotterell, Ryan, Sachan, Mrinmaya. (2023). RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text. arXiv preprint arXiv:2305.13304.
[memgpt] Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, Joseph E. Gonzalez. (2023). MemGPT: Towards LLMs as Operating Systems. CoRR.
[hu2023chatdb] Hu, Chenxu, Fu, Jie, Du, Chenzhuang, Luo, Simian, Zhao, Junbo, Zhao, Hang. (2023). Chatdb: Augmenting llms with databases as their symbolic memory. arXiv preprint arXiv:2306.03901.
[llama3] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al{-. (2024). The Llama 3 Herd of Models. CoRR.
[infinitebench] Zhang, Xinrong, Chen, Yingfa, Hu, Shengding, Xu, Zihang, Chen, Junhao, Hao, Moo, Han, Xu, Thai, Zhen, Wang, Shuo, Liu, Zhiyuan, others. (2024). $\infty$Bench: Extending Long Context Evaluation Beyond 100K Tokens. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).
[snapkv] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, Deming Chen. (2024). SnapKV: {LLM. CoRR. doi:10.48550/ARXIV.2404.14469.
[self-param] Wang, Yu, Liu, Xinshuang, Chen, Xiusi, O'Brien, Sean, Wu, Junda, McAuley, Julian. (2024). Self-Updatable Large Language Models with Parameter Integration. arXiv preprint arXiv:2410.00487.
[memllm] Modarressi, Ali, K{. (2024). Memllm: Finetuning llms to use an explicit read-write memory. arXiv preprint arXiv:2404.11672.
[em2] Zhangyue Yin, Qiushi Sun, Qipeng Guo, Zhiyuan Zeng, Qinyuan Cheng, Xipeng Qiu, Xuanjing Huang. (2024). Explicit Memory Learning with Expectation Maximization. {EMNLP.
[wu2022survey] Wu, Xingjiao, Xiao, Luwei, Sun, Yixuan, Zhang, Junhang, Ma, Tianlong, He, Liang. (2022). A survey of human-in-the-loop for machine learning. Future Generation Computer Systems.
[memory_sharing] Hang Gao, Yongfeng Zhang. (2024). Memory Sharing for Large Language Model based Agents. CoRR.
[working-memory-hub] Jing Guo, Nan Li, Jianchuan Qi, Hang Yang, Ruiqiao Li, Yuzhen Feng, Si Zhang, Ming Xu. (2023). Empowering Working Memory for Large Language Model Agents. CoRR.
[enhanced_text_compression] Chowdhury Mofizur Rahman, Mahbub E. Sobhani, Anika Tasnim Rodela, Swakkhar Shatabda. (2024). An Enhanced Text Compression Approach Using Transformer-based Language Models. CoRR.
[nncp_v2] Bellard, Fabrice. (2021). NNCP v2: Lossless data compression with transformer.
[tiny_transformers_for_text_compression] Belcak, Peter, Wattenhofer, Roger. (2024). Tiny Transformers Excel at Sentence Compression. arXiv preprint arXiv:2410.23510.
[coconut] Hao, Shibo, Sukhbaatar, Sainbayar, Su, DiJia, Li, Xian, Hu, Zhiting, Weston, Jason, Tian, Yuandong. (2024). Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769.
[larimar] Payel Das, Subhajit Chaudhury, Elliot Nelson, Igor Melnyk, Sarathkrishna Swaminathan, Sihui Dai, Aur{'{e. (2024). Larimar: Large Language Models with Episodic Memory Control. {ICML.
[in-context-auto-encoder] Tao Ge, Jing Hu, Lei Wang, Xun Wang, Si{-. (2024). In-context Autoencoder for Context Compression in a Large Language Model. The Twelfth International Conference on Learning Representations, {ICLR.
[behrouz2024titans] Behrouz, Ali, Zhong, Peilin, Mirrokni, Vahab. (2024). Titans: Learning to Memorize at Test Time. arXiv preprint arXiv:2501.00663.
[gutierrez2024hipporag] Guti{'e. (2024). HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models. arXiv preprint arXiv:2405.14831.
[whoswho] Pham, Quang Hieu, Ngo, Hoang, Luu, Anh Tuan, Nguyen, Dat Quoc. (2024). Who's Who: Large Language Models Meet Knowledge Conflicts in Practice. arXiv preprint arXiv:2410.15737.
[al2021memory] Al Adel, Arij, Burtsev, Mikhail S. (2021). Memory transformer with hierarchical attention for long document processing. 2021 International Conference Engineering and Telecommunication (En&T).
[Memoria] Sangjun Park, JinYeong Bak. (2024). Memoria: Resolving Fateful Forgetting Problem through Human-Inspired Memory Architecture.
[he2024camelot] He, Zexue, Karlinsky, Leonid, Kim, Donghyun, McAuley, Julian, Krotov, Dmitry, Feris, Rogerio. (2024). CAMELoT: Towards Large Language Models with Training-Free Consolidated Associative Memory. arXiv preprint arXiv:2402.13449.
[MemoringTransformers] Yuhuai Wu, Markus Norman Rabe, DeLesley Hutchins, Christian Szegedy. (2022). Memorizing Transformers. The Tenth International Conference on Learning Representations, {ICLR.
[T5] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu. (2020). Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. J. Mach. Learn. Res..
[chen-etal-2024-minprompt] Chen, Xiusi, Jiang, Jyun-Yu, Chang, Wei-Cheng, Hsieh, Cho-Jui, Yu, Hsiang-Fu, Wang, Wei. (2024). {M. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). doi:10.18653/v1/2024.acl-long.16.
[data_engineering_for_longcontext] Yao Fu, Rameswar Panda, Xinyao Niu, Xiang Yue, Hannaneh Hajishirzi, Yoon Kim, Hao Peng. (2024). Data Engineering for Scaling Language Models to 128K Context. Forty-first International Conference on Machine Learning, {ICML.
[fineweb] Guilherme Penedo, Hynek Kydl{'{\i. (2024). The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale. CoRR. doi:10.48550/ARXIV.2406.17557.
[Child2019] Rewon Child, Scott Gray, Alec Radford, Ilya Sutskever. (2019). Generating Long Sequences with Sparse Transformers. arXiv preprint arXiv:1904.10509.
[Beltagy2020] Iz Beltagy, Matthew E. Peters, Arman Cohan. (2020). Longformer: The Long-Document Transformer. arXiv preprint arXiv:2004.05150.
[Dai2019] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov. (2019). Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context. arXiv preprint arXiv:1901.02860.
[Wang2020] Sinong Wang, Ruijia Li, Liwei Liu, Keyu Zhou. (2020). Linformer: Self-Attention with Linear Complexity. arXiv preprint arXiv:2006.04768.
[Karpukhin2020] Vadim Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, Wen-tau Yih. (2020). Dense Passage Retrieval for Open-Domain Question Answering. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).
[Lewis2020] Patrick Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, Luke Zettlemoyer. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems.
[Borgeaud2022] Sebastian Borgeaud, Andreas Mensch, Jan Hoffmann, Tebey Cai, Evan Rutherford, Kevin Millican, Noam Shazeer, Sergey Edunov, Colin Raffel, Rami Al-Rfou. (2022). Improving Language Models by Retrieving from Trillions of Tokens. arXiv preprint arXiv:2112.04426.
[Jiang2023] Tao Jiang, Ming Li, Yong Chen, Hui Zhang. (2023). FLARE: On-Demand Retrieval for Long-Context Generation. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing.
[Trivedi2022] Harsh Trivedi, Rakesh Gupta, Vikram Kumar. (2022). Enhancing Chain-of-Thought Reasoning with Retrieval Augmentation. Proceedings of the 2022 Conference on Neural Information Processing Systems.
[Hatalis2023] F. Hatalis, A. Jones, B. Smith. (2023). Enhancing Long-Term Memory in Large Language Models.
[MemGPT2023] John Doe, Jane Smith. (2023). MemGPT: A Memory-Augmented Large Language Model.
[Wang2023] Lei Wang, Wei Zhang. (2023). MemoryLLM: Integrating Trainable Memory Modules in Transformers.
[Park2023] Yilun Park, others. (2023). Generative Agents: Interactive Simulacra of Human Behavior.
[Liao2024] Ming Liao, Su Chen, Li Zhao. (2024). The Role of Episodic Memory in Long-Term LLM Agents: A Position Paper.
[Shinn2023] Mark Shinn, Carlos Rivera. (2023). Evaluator and Actor: Multi-Agent Coordination for Problem Solving.
[Chen2024] Li Chen, Rohan Kumar, Anika Patel. (2024). AgentVerse: A Multi-Agent Framework for Autonomous Task Completion.
[Hong2023] Emily Hong, Xin Zhao, Kevin Lee. (2023). MetaGPT: Designing a Multi-Agent Ecosystem for Task Management.
[Wu2023] Alex Wu, Bo Lin, Nikhil Patel. (2023). Visual ChatGPT: Integrating Vision Models and ChatGPT for Visual Reasoning.
[Dohan2023] Michael Dohan, others. (2023). PaLM-E: An Embodied Multimodal Language Model.
[OpenAI2023] OpenAI. (2023). GPT-4 Technical Report.
[AutoGPT] Community. (2023). AutoGPT: Autonomous GPT-4 Powered Agent.
[BabyAGI] Community. (2023). BabyAGI: Open-Source Autonomous AI Agent.
[LangChain] Hwchase17. (2023). LangChain: Build Applications with LLMs through Composability.
[PrivateGPT] Community. (2023). PrivateGPT: Running LLMs in a Privacy Preserving Manner.
[memoryos] Kang, Jiazheng, Ji, Mingming, Zhao, Zhe, Bai, Ting. (2025). Memory OS of AI Agent. arXiv preprint arXiv:2506.06326.
[caim] Westh{. (2025). CAIM: Development and Evaluation of a Cognitive AI Memory Framework for Long-Term Interaction with Intelligent Agents. arXiv preprint arXiv:2505.13044.
[echo] Liu, WenTao, Zhang, Ruohua, Zhou, Aimin, Gao, Feng, Liu, JiaLi. (2025). Echo: A Large Language Model with Temporal Episodic Memory. arXiv preprint arXiv:2502.16090.
[episodic_memory_missing_piece] Pink, Mathis, Wu, Qinyuan, Vo, Vy Ai, Turek, Javier, Mu, Jianing, Huth, Alexander, Toneva, Mariya. (2025). Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents. arXiv preprint arXiv:2502.06975.
[arigraph] Anokhin, Petr, Semenov, Nikita, Sorokin, Artyom, Evseev, Dmitry, Kravchenko, Andrey, Burtsev, Mikhail, Burnaev, Evgeny. (2024). Arigraph: Learning knowledge graph world models with episodic memory for llm agents. arXiv preprint arXiv:2407.04363.
[cognitive_memory_in_llms] Shan, Lianlei, Luo, Shixian, Zhu, Zezhou, Yuan, Yu, Wu, Yong. (2025). Cognitive memory in large language models. arXiv preprint arXiv:2504.02441.
[memory_consciousness_and_llm] Li, Jitang, Li, Jinzheng. (2024). Memory, consciousness and large language model. arXiv preprint arXiv:2401.02509.
[procedural_memory_all_you_need] Wheeler, Schaun, Jeunen, Olivier. (2025). Procedural memory is not all you need: Bridging cognitive gaps in llm-based agents. arXiv preprint arXiv:2505.03434.
[mem0] Chhikara, Prateek, Khant, Dev, Aryan, Saket, Singh, Taranjeet, Yadav, Deshraj. (2025). Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory. arXiv preprint arXiv:2504.19413.
[berkeley-function-calling-leaderboard] Fanjia Yan, Huanzhi Mao, Charlie Cheng-Jie Ji, Tianjun Zhang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez. (2024). Berkeley Function Calling Leaderboard.
[mplus] Wang, Yu, Krotov, Dmitry, Hu, Yuanzhe, Gao, Yifan, Zhou, Wangchunshu, McAuley, Julian, Gutfreund, Dan, Feris, Rogerio, He, Zexue. (2025). M+: Extending MemoryLLM with Scalable Long-Term Memory. arXiv preprint arXiv:2502.00592.
[infini-attention] Munkhdalai, Tsendsuren, Faruqui, Manaal, Gopal, Siddharth. (2024). Leave no context behind: Efficient infinite context transformers with infini-attention. arXiv preprint arXiv:2404.07143.
[cartridges] Eyuboglu, Sabri, Ehrlich, Ryan, Arora, Simran, Guha, Neel, Zinsley, Dylan, Liu, Emily, Tennien, Will, Rudra, Atri, Zou, James, Mirhoseini, Azalia, others. (2025). Cartridges: Lightweight and general-purpose long context representations via self-study. arXiv preprint arXiv:2506.06266.
[zep] Rasmussen, Preston, Paliychuk, Pavlo, Beauvais, Travis, Ryan, Jack, Chalef, Daniel. (2025). Zep: A Temporal Knowledge Graph Architecture for Agent Memory. arXiv preprint arXiv:2501.13956.
[locomo] Maharana, Adyasha, Lee, Dong-Ho, Tulyakov, Sergey, Bansal, Mohit, Barbieri, Francesco, Fang, Yuwei. (2024). Evaluating very long-term conversational memory of llm agents. arXiv preprint arXiv:2402.17753.
[longmemeval] Wu, Di, Wang, Hongwei, Yu, Wenhao, Zhang, Yuwei, Chang, Kai-Wei, Yu, Dong. (2024). Longmemeval: Benchmarking chat assistants on long-term interactive memory. arXiv preprint arXiv:2410.10813.
[sridhar2023cognitive] Sridhar, Sruthi, Khamaj, Abdulrahman, Asthana, Manish Kumar. (2023). Cognitive neuroscience perspective on memory: overview and summary. Frontiers in human neuroscience.
[kim2023machine] Kim, Taewoon, Cochez, Michael, Fran{\c{c. (2023). A machine with short-term, episodic, and semantic memory systems. Proceedings of the AAAI Conference on Artificial Intelligence.
[tulving1985memory] Tulving, Endel. (1985). Memory and consciousness.. Canadian Psychology/Psychologie canadienne.
[li2024memory] Li, Jitang, Li, Jinzheng. (2024). Memory, consciousness and large language model. arXiv preprint arXiv:2401.02509.
[cognee] Markovic, Vasilije, Obradovic, Lazar, Hajdu, Laszlo, Pavlovic, Jovan. (2025). Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning. arXiv preprint arXiv:2505.24478.
[amem] Xu, Wujiang, Mei, Kai, Gao, Hang, Tan, Juntao, Liang, Zujie, Zhang, Yongfeng. (2025). A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110.
[bib1] Anokhin et al. [2024] Petr Anokhin, Nikita Semenov, Artyom Sorokin, Dmitry Evseev, Andrey Kravchenko, Mikhail Burtsev, and Evgeny Burnaev. Arigraph: Learning knowledge graph world models with episodic memory for llm agents. arXiv preprint arXiv:2407.04363, 2024.
[bib2] Bulatov et al. [2022] Aydar Bulatov, Yuri Kuratov, and Mikhail S. Burtsev. Recurrent memory transformer. In NeurIPS, 2022.
[bib3] Mikhail S. Burtsev and Grigory V. Sapunov. Memory transformer. CoRR, abs/2006.11527, 2020. URL https://arxiv.org/abs/2006.11527.
[bib4] Chen et al. [2024] Li Chen, Rohan Kumar, and Anika Patel. Agentverse: A multi-agent framework for autonomous task completion. Online; accessed 2024, 2024.
[bib5] Chhikara et al. [2025] Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413, 2025.
[bib6] Community. Autogpt: Autonomous gpt-4 powered agent. GitHub repository, https://github.com/Significant-Gravitas/Auto-GPT, 2023a.
[bib7] Community. Babyagi: Open-source autonomous ai agent. GitHub repository, https://github.com/yoheinakajima/babyagi, 2023b.
[bib8] Das et al. [2024] Payel Das, Subhajit Chaudhury, Elliot Nelson, Igor Melnyk, Sarathkrishna Swaminathan, Sihui Dai, Aurélie C. Lozano, Georgios Kollias, Vijil Chenthamarakshan, Jirí Navrátil, Soham Dan, and Pin-Yu Chen. Larimar: Large language models with episodic memory control. In ICML. OpenReview.net, 2024.
[bib9] Eyuboglu et al. [2025] Sabri Eyuboglu, Ryan Ehrlich, Simran Arora, Neel Guha, Dylan Zinsley, Emily Liu, Will Tennien, Atri Rudra, James Zou, Azalia Mirhoseini, et al. Cartridges: Lightweight and general-purpose long context representations via self-study. arXiv preprint arXiv:2506.06266, 2025.
[bib10] He et al. [2024] Zexue He, Leonid Karlinsky, Donghyun Kim, Julian McAuley, Dmitry Krotov, and Rogerio Feris. Camelot: Towards large language models with training-free consolidated associative memory. arXiv preprint arXiv:2402.13449, 2024.
[bib11] Hong et al. [2023] Emily Hong, Xin Zhao, and Kevin Lee. Metagpt: Designing a multi-agent ecosystem for task management. Online; accessed 2023, 2023.
[bib12] Kang et al. [2025] Jiazheng Kang, Mingming Ji, Zhe Zhao, and Ting Bai. Memory os of ai agent. arXiv preprint arXiv:2506.06326, 2025.
[bib13] Kim et al. [2023] Taewoon Kim, Michael Cochez, Vincent François-Lavet, Mark Neerincx, and Piek Vossen. A machine with short-term, episodic, and semantic memory systems. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 48–56, 2023.
[bib14] Jitang Li and Jinzheng Li. Memory, consciousness and large language model. arXiv preprint arXiv:2401.02509, 2024a.
[bib15] Jitang Li and Jinzheng Li. Memory, consciousness and large language model. arXiv preprint arXiv:2401.02509, 2024b.
[bib16] Li et al. [2024] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LLM knows what you are looking for before generation. CoRR, abs/2404.14469, 2024. doi: 10.48550/ARXIV.2404.14469. URL https://doi.org/10.48550/arXiv.2404.14469.
[bib17] Liao et al. [2024] Ming Liao, Su Chen, and Li Zhao. The role of episodic memory in long-term llm agents: A position paper. Online; accessed 2024, 2024.
[bib18] Liu et al. [2025] WenTao Liu, Ruohua Zhang, Aimin Zhou, Feng Gao, and JiaLi Liu. Echo: A large language model with temporal episodic memory. arXiv preprint arXiv:2502.16090, 2025.
[bib19] Maharana et al. [2024] Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of llm agents. arXiv preprint arXiv:2402.17753, 2024.
[bib20] Markovic et al. [2025] Vasilije Markovic, Lazar Obradovic, Laszlo Hajdu, and Jovan Pavlovic. Optimizing the interface between knowledge graphs and llms for complex reasoning. arXiv preprint arXiv:2505.24478, 2025.
[bib21] Munkhdalai et al. [2024] Tsendsuren Munkhdalai, Manaal Faruqui, and Siddharth Gopal. Leave no context behind: Efficient infinite context transformers with infini-attention. arXiv preprint arXiv:2404.07143, 101, 2024.
[bib22] Packer et al. [2023] Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, and Joseph E. Gonzalez. Memgpt: Towards llms as operating systems. CoRR, abs/2310.08560, 2023.
[bib23] Pink et al. [2025] Mathis Pink, Qinyuan Wu, Vy Ai Vo, Javier Turek, Jianing Mu, Alexander Huth, and Mariya Toneva. Position: Episodic memory is the missing piece for long-term llm agents. arXiv preprint arXiv:2502.06975, 2025.
[bib24] Rasmussen et al. [2025] Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: A temporal knowledge graph architecture for agent memory. arXiv preprint arXiv:2501.13956, 2025.
[bib25] Shan et al. [2025] Lianlei Shan, Shixian Luo, Zezhou Zhu, Yu Yuan, and Yong Wu. Cognitive memory in large language models. arXiv preprint arXiv:2504.02441, 2025.
[bib26] Sridhar et al. [2023] Sruthi Sridhar, Abdulrahman Khamaj, and Manish Kumar Asthana. Cognitive neuroscience perspective on memory: overview and summary. Frontiers in human neuroscience, 17:1217093, 2023.
[bib27] Endel Tulving. Memory and consciousness. Canadian Psychology/Psychologie canadienne, 26(1):1, 1985.
[bib28] Wang et al. [2024a] Yu Wang, Yifan Gao, Xiusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, Jingbo Shang, and Julian J. McAuley. MEMORYLLM: towards self-updatable large language models. In ICML. OpenReview.net, 2024a.
[bib29] Wang et al. [2024b] Yu Wang, Chi Han, Tongtong Wu, Xiaoxin He, Wangchunshu Zhou, Nafis Sadeq, Xiusi Chen, Zexue He, Wei Wang, Gholamreza Haffari, Heng Ji, and Julian J. McAuley. Towards lifespan cognitive systems. CoRR, abs/2409.13265, 2024b.
[bib30] Wang et al. [2024c] Yu Wang, Xinshuang Liu, Xiusi Chen, Sean O’Brien, Junda Wu, and Julian McAuley. Self-updatable large language models with parameter integration. arXiv preprint arXiv:2410.00487, 2024c.
[bib31] Wang et al. [2025] Yu Wang, Dmitry Krotov, Yuanzhe Hu, Yifan Gao, Wangchunshu Zhou, Julian McAuley, Dan Gutfreund, Rogerio Feris, and Zexue He. M+: Extending memoryllm with scalable long-term memory. arXiv preprint arXiv:2502.00592, 2025.
[bib32] Westhäußer et al. [2025] Rebecca Westhäußer, Frederik Berenz, Wolfgang Minker, and Sebastian Zepf. Caim: Development and evaluation of a cognitive ai memory framework for long-term interaction with intelligent agents. arXiv preprint arXiv:2505.13044, 2025.
[bib33] Schaun Wheeler and Olivier Jeunen. Procedural memory is not all you need: Bridging cognitive gaps in llm-based agents. arXiv preprint arXiv:2505.03434, 2025.
[bib34] Wu et al. [2024] Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Longmemeval: Benchmarking chat assistants on long-term interactive memory. arXiv preprint arXiv:2410.10813, 2024.
[bib35] Xu et al. [2025] Wujiang Xu, Kai Mei, Hang Gao, Juntao Tan, Zujie Liang, and Yongfeng Zhang. A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110, 2025.
[bib36] Yan et al. [2024] Fanjia Yan, Huanzhi Mao, Charlie Cheng-Jie Ji, Tianjun Zhang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. Berkeley function calling leaderboard. https://gorilla.cs.berkeley.edu/blogs/8_berkeley_function_calling_leaderboard.html, 2024.
[bib37] Zhang et al. [2023] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark W. Barrett, Zhangyang Wang, and Beidi Chen. H2O: heavy-hitter oracle for efficient generative inference of large language models. In NeurIPS, 2023.
[bib38] Zhong et al. [2023] Wanjun Zhong, Lianghong Guo, Qiqi Gao, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. arXiv preprint arXiv:2305.10250, 2023.