// the find
BAI-LAB/MemoryOS
[EMNLP 2025 Oral] MemoryOS is designed to provide a memory operating system for personalized AI agents.
MemoryOS is a hierarchical memory management library for LLM-based agents, organizing conversation history into short-term, mid-term, and long-term layers — analogous to CPU registers, RAM, and disk. It has a peer-reviewed paper (EMNLP 2025) with benchmark numbers behind it, and ships as a PyPI package plus an MCP server for dropping into Claude Desktop or Cline. For anyone building agents that need to remember users across sessions without just stuffing everything into the context window, this is a reasonable starting point.
The three-tier memory model is genuinely thought through — mid-term 'heat' scoring means frequently-revisited topics bubble up to long-term storage rather than everything being treated equally. The MCP server variant means you can add persistent memory to any MCP-compatible client without writing integration code. The LoCoMo benchmark numbers are reproducible (eval scripts are in the repo) which is more than most memory libraries offer. Docker image and ChromaDB variant are both there, so you're not locked into flat-file storage from day one.
The codebase is duplicated across four directories (memoryos-pypi, memoryos-mcp, memoryos-chromadb, memoryos-playground) with no shared core — a bug fix in one won't reach the others, and the variants are already visibly diverging. All LLM calls go through the OpenAI interface, so Anthropic support means pointing at an OpenAI-compatible proxy rather than the native SDK — fine for now but fragile if you want tool use or streaming from Claude directly. The heat-threshold parameters (mid_term_heat_threshold, similarity_threshold) have no guidance on how to tune them for different conversation volumes; the defaults feel like they were chosen for the benchmark dataset. Short-term capacity defaults to 7 QA pairs, which is small enough that a normal onboarding conversation will flush before any user facts reach long-term memory.