MIRIX: Modular Multi-Agent Memory System Boosts Long-Term LLM Capabilities

Long-term memory remains a key hurdle for large language model (LLM) agents aiming to deliver consistent, personalized interactions over time. Most existing LLM agents operate without persistent memory, limiting their ability to recall past user information beyond isolated prompts. Addressing this challenge, MIRIX AI has developed MIRIX, a modular multi-agent memory system designed to enhance long-term memory and reasoning in LLM-based agents.

How MIRIX Advances Memory in LLM Agents

Unlike conventional memory approaches that rely mainly on text truncation or simple retrieval, MIRIX introduces a structured, multi-agent architecture featuring six specialized memory components. Each component is managed by an independent Memory Manager, coordinated by a Meta Memory Manager that intelligently routes queries and manages hierarchical storage.

The six core memory types include:

• Core Memory: Stores persistent agent and user profiles, including persona details (tone, behavior) and human data such as preferences and relationships.
• Episodic Memory: Captures time-stamped events and user interactions with structured attributes like event summaries and participants.
• Semantic Memory: Encodes abstract knowledge such as named entities and knowledge graphs.
• Procedural Memory: Maintains detailed workflows and task sequences, often stored in JSON for easy manipulation.
• Resource Memory: References external files including documents, images, and audio for contextual continuity.
• Knowledge Vault: Holds sensitive facts such as credentials and API keys with strict access controls.

This sophisticated layering allows MIRIX to maintain rich, multimodal memory—including visual input—and deliver context-aware, personalized responses.

A key innovation is the Active Retrieval pipeline: upon receiving user input, MIRIX autonomously infers the relevant topic, retrieves matching memory entries from all components using strategies like embedding, BM25, and string matching, and injects the contextual data back into the system prompt. This method reduces dependence on the static knowledge stored within an LLM’s parameters and significantly strengthens response grounding.

Deployed as a cross-platform assistant application built with React-Electron and Uvicorn, MIRIX captures screenshots every 1.5 seconds to track visual context, retaining only unique images for memory updates. Visual data is streamed through the Gemini API, enabling memory refreshes with under 5 seconds latency during active use. Users engage with MIRIX via a chat interface that transparently renders semantic and procedural memories, making it possible to audit and interact with the agent’s knowledge base.

Rigorous evaluation highlights MIRIX’s efficacy:

• ScreenshotVQA Benchmark: Tasked with answering questions from high-res screenshots over long periods, MIRIX outperforms existing retrieval-augmented generation models by 35% in accuracy while drastically cutting storage needs compared to text-heavy methods.
• LOCOMO Conversation Benchmark: MIRIX achieves 85.38% average accuracy for long-form conversational memory, surpassing strong open-source counterparts by over 8 points and closely approaching the full-context upper bound.

Designed with scalability in mind, MIRIX supports deployment on lightweight AI wearables, such as smart glasses, by allowing hybrid on-device and cloud memory management. Practical applications include real-time meeting summaries, location-aware recall, and dynamic modeling of user habits.

Notably, MIRIX envisions a Memory Marketplace—a decentralized platform enabling secure, privacy-conscious memory sharing and monetization among users. This marketplace emphasizes fine-grained privacy controls, end-to-end encryption, and decentralized storage to empower user data sovereignty.

By integrating modular, multimodal memory layers with a collaborative multi-agent framework, MIRIX pushes the frontier for persistent, context-rich LLM agents. Its flexible architecture and demonstrated performance gains mark an important milestone in AI memory systems.

As AI agents strive to move beyond single-session interactions toward continuous personalization, solutions like MIRIX highlight the critical role of structured long-term memory in achieving truly intelligent assistance.