MIRIX: Multi-Agent Memory System for LLM-Based Agents

Although memory capabilities of AI agents are gaining increasing attention,existing solutions remain fundamentally limited. Most rely on flat, narrowlyscoped memory components, constraining their ability to personalize, abstract,and reliably recall user-specific information over time. To this end, weintroduce MIRIX, a modular, multi-agent memory system that redefines the futureof AI memory by solving the field's most critical challenge: enabling languagemodels to truly remember. Unlike prior approaches, MIRIX transcends text toembrace rich visual and multimodal experiences, making memory genuinely usefulin real-world scenarios. MIRIX consists of six distinct, carefully structuredmemory types: Core, Episodic, Semantic, Procedural, Resource Memory, andKnowledge Vault, coupled with a multi-agent framework that dynamically controlsand coordinates updates and retrieval. This design enables agents to persist,reason over, and accurately retrieve diverse, long-term user data at scale. Wevalidate MIRIX in two demanding settings. First, on ScreenshotVQA, achallenging multimodal benchmark comprising nearly 20,000 high-resolutioncomputer screenshots per sequence, requiring deep contextual understanding andwhere no existing memory systems can be applied, MIRIX achieves 35% higheraccuracy than the RAG baseline while reducing storage requirements by 99.9%.Second, on LOCOMO, a long-form conversation benchmark with single-modal textualinput, MIRIX attains state-of-the-art performance of 85.4%, far surpassingexisting baselines. These results show that MIRIX sets a new performancestandard for memory-augmented LLM agents. To allow users to experience ourmemory system, we provide a packaged application powered by MIRIX. It monitorsthe screen in real time, builds a personalized memory base, and offersintuitive visualization and secure local storage to ensure privacy.