EverMemOS is an open-source, enterprise-grade intelligent memory system designed to equip AI agents with persistent, context-aware long-term memory. It moves beyond traditional "look-back" databases, enabling AI to understand the meaning behind memories and use them for foresight, resulting in more coherent, personalized, and intelligent conversations. This system is ideal for developers building advanced AI agents that require deep contextual understanding and continuous learning.

How It Works

The system operates via two core tracks: memory construction and memory perception. Memory construction builds structured, retrievable long-term memory from raw conversation data using atomic "MemCells," integrating related fragments into multi-level memories organized by theme and storyline. Memory perception then intelligently recalls relevant memories through multi-round reasoning and fusion. It employs hybrid retrieval (Reciprocal Rank Fusion), intelligent reranking, and LLM-guided multi-round recall to achieve precise contextual awareness, allowing AI to proactively leverage past information for current decision-making.

Quick Start & Requirements

Installation involves cloning the repository, starting dependency services (MongoDB, Elasticsearch, Milvus, Redis) via Docker Compose, and installing project dependencies using uv sync. Prerequisites include Python 3.10+, Docker 20.10+, and at least 4GB of RAM for database services. LLM API keys (e.g., LLM_API_KEY, DEEPINFRA_API_KEY) are necessary for core functionalities. Demos are available for quick testing after setup.

Highlighted Details

• Achieved 92.3% reasoning accuracy on the LoCoMo benchmark.
• Features "Coherent Narrative" by automatically linking conversation fragments into thematic context.
• Enables "Evidence-Based Perception" by proactively capturing deep connections between memories and tasks.
• Maintains "Living Profiles" that dynamically update user understanding with each interaction.

Maintenance & Community

The project is maintained by the EverMemOS Team. While aiming to build a vibrant open-source community, specific community channels (like Discord or Slack) are not detailed in the provided README.

Licensing & Compatibility

This project is licensed under the Apache License 2.0, permitting free use, modification, and distribution with standard attribution and change notification requirements. It is generally compatible with commercial use and closed-source linking.

Limitations & Caveats

A technical paper is currently "Coming Soon." Core functionalities heavily rely on external LLM API keys. The system requires significant RAM for its database services.