MemOS
Self-evolving memory OS for LLM & AI Agents: ultra-persistent memory, hybrid-retrieval, and cross-task skill reuse, with 35.24% token savings
MemOS gives AI assistants and chatbots a persistent memory so they can remember past conversations and decisions across sessions, instead of starting completely fresh each time.
MemOS is a memory system for AI agents and large language models. The basic problem it addresses is that most AI chat systems start fresh with each conversation: they have no recollection of what was discussed before unless you paste it back in manually. MemOS gives AI agents a structured, persistent memory that survives across sessions, so the agent can recall earlier conversations, build on past decisions, and adapt its behavior over time.
The system organizes memory into different layers. Short-term interaction traces sit at one level, learned patterns and policies at another, and a broader model of the world at a third. There is also a layer for reusable skills, things the agent has figured out how to do that it can apply again in new situations. These layers are not just storage: the system evolves them over time based on feedback, adding, correcting, and replacing memories as needed.
For retrieval, MemOS uses a hybrid approach that combines traditional keyword search with vector-based semantic search, so it can find relevant memories whether the query matches the exact phrasing or just the general meaning. The project claims this reduces the number of tokens needed in AI prompts by about 35%, because you can retrieve the specific memory that matters rather than stuffing the entire history into every request.
MemOS ships as two main options: a self-hosted local plugin and a cloud service. The local plugin stores everything on the user's device using SQLite, with no external dependency. The cloud version adds multi-agent memory sharing, where multiple AI agents working under the same user ID can share a common memory pool. Both variants expose a single API for adding, retrieving, editing, and deleting memories.
The project is designed to integrate with existing AI agent frameworks. Plugins are available for Hermes Agent and OpenClaw. It supports text, images, tool usage history, and knowledge bases built from documents or URLs. MemOS is accompanied by a research paper and is open source under a license the README links to its documentation site for full details.
Where it fits
- Give a customer service chatbot memory of past conversations so users do not have to repeat themselves in every new session.
- Build a personal AI assistant that remembers your preferences and past decisions across separate chat sessions.
- Reduce AI prompt token costs by storing conversation history in MemOS and retrieving only the relevant parts each time.
- Enable multiple AI agents working on the same project to share a common memory pool through the cloud version.