MIRIX: A Modular Multi-Agent Memory System for Enhanced Long-Term Reasoning and Personalization in LLM-Based Agents

The article introduces MIRIX, a modular multi-agent memory system designed to enhance long-term reasoning and personalization in large language model (LLM) agents. Unlike traditional stateless models, MIRIX incorporates structured, multimodal memory componentsincluding visual inputsmanaged by specialized Memory Managers, enabling persistent storage of user profiles, agent personas, and episodic interactions over time. This architecture addresses a critical gap in LLM capabilities by allowing agents to recall and reason over user-specific information across sessions, thereby improving contextual continuity and personalized experiences. MIRIX's core architecture features six distinct memory modules, such as Core Memory