MemoryAtlas

An OS-inspired memory layer for personalized AI agents that organizes user memory into short-, mid-, and long-term tiers and migrates entries between them via heat-based promotion/demotion. Four core modules: Storage, Updating, Retrieval, and Generation. Published as an EMNLP 2025 Oral.

Storage
Segmented-page storage with a heat-based updating mechanism that promotes/demotes memory across short / mid / long-term tiers; ChromaDB vector database supported as an alternative backend.
Retrieval
Retrieves relevant context from short-term history, mid-term segments, and long-term user profile/knowledge base; supports BGE-M3 and all-MiniLM-L6-v2 embeddings; configurable similarity_threshold and retrieval queue capacity.
Self-host
Self-host: moderate
License
Apache-2.0
Pricing
Open source (Apache-2.0), free to self-host · Free / OSS
GitHub stars
1,551
Last release
2025-07-18
Last commit
2026-07-07
First catalogued
2026-06-28

Strengths

  • Peer-reviewed (EMNLP 2025 Oral); authors report a 49.11% F1 gain on LoCoMo (self-reported, not independently reproduced — route to harvest-benchmarks)
  • OS-style tiered memory with heat-based promotion/demotion
  • First-party MCP server (local-stdio) for Claude Desktop, Cline, and Cursor
  • Python SDK on PyPI (memoryos-pro); Docker image available
  • Permissive Apache-2.0 license

Watch out

  • Name collides with the separate MemTensor MemOS project — they are different systems
  • No versioned GitHub release since V1.2 (2025-07-18); repo does see periodic bug-fix commits (e.g. a thread-safety fix to the MCP/playground stores landed 2026-07-06/07) but maintenance cadence is research-project-paced, not release-driven
  • Python SDK at v0.1.0 (single PyPI release as of 2025-07-09); API surface may change

Best for

  • Personalized conversational agents needing tiered long-term user memory
  • Claude Desktop / Cursor / Cline setups wanting plug-in memory via MCP

How it integrates

Benchmark results

No sourced results yet.

Sources

Last verified 2026-08-01 · updated by refresh-framework-cards