Awesome AI Agents › AI Agent Operating System

BAI-LAB/MemoryOS

⭐ 1575 Python repository created 2025-05-30

MemoryOS is a memory management layer for personalized AI agents, designed so that assistants stay coherent and context-aware across long conversations. Its design borrows from operating system memory management: memory is held in a hierarchical store split into short-term, mid-term and long-term persona memory, and four core modules handle storage, updating, retrieval and generation. User profiles and accumulated knowledge are updated automatically as dialogue proceeds. The architecture is described as plug and play, so storage engines, update strategies and retrieval algorithms can be swapped as separate modules. The project reports state of the art results on the LoCoMo long-term memory benchmark, with average gains of about 49 percent in F1 and 46 percent in BLEU-1 over the compared baseline, and the evaluation scripts for reproducing those numbers are published. Alongside the Python package, the maintainers ship MemoryOS-MCP, a Model Context Protocol server that injects long-term memory into agent clients such as Claude Desktop by exposing memory operations as callable tools, and a Playground platform for interactive use. Deployment options include a PyPI package, a ChromaDB-backed variant for vector storage, and Docker. A range of embedding models is supported, including BGE-M3 and Qwen3, and the memory layer works with many LLM providers such as OpenAI, Deepseek and Qwen, including reasoning models like Deepseek-R1. Configuration parameters such as a similarity threshold are documented on a dedicated documentation site. The work is published as a paper on arXiv and was accepted at the EMNLP 2025 main conference, and it belongs to a wider research line on AI memory that also includes a survey and a deep search system built on experiential memory. Licensed under Apache 2.0.

https://github.com/BAI-LAB/MemoryOS

memoryllm-agentsmcpretrievalpythonpersonalization