MemoryAtlas

Rejects background auto-extraction on principle — the project's position is that with proxy memory 'the AI doesn't know what it remembers and has no say in what gets recorded.' Instead the agent itself authors, classifies, and writes its own recall trigger for every memory through seven first-person MCP tools. A Python/FastAPI backend stores memories as a node–memory–edge–path graph addressed by human-readable URIs (core://agent/identity), with an automatic version-chain snapshot on every AI write and a React dashboard offering diff and rollback.

Storage
SQLite by default (a single file, no external service) or PostgreSQL for multi-device Docker deployments. There is no vector database — content is stored and queried structurally rather than embedded.
Retrieval
Three paths, none of them semantic: direct URI reads (domain://path, plus system entry points like system://boot and system://recent), substring keyword search, and Aho-Corasick trigger-word matching that auto-surfaces cross-node disclosure conditions instead of cosine-similarity ranking.
Self-host
Self-host: trivial
License
MIT
Pricing
Free and open source with no paid tier of any kind. The only hosted surface is a read-only public demo with sample data, so self-hosting is effectively required for real use. · Free / OSS
GitHub stars
1,278
Last release
2026-05-31
Last commit
2026-07-20
First catalogued
2026-07-25

Strengths

  • Genuinely novel data model — agent-authored, version-controlled, human-auditable graph memory with rollback rather than another vector-RAG wrapper
  • A full 823-line English README exists alongside the Chinese original, plus a translated system prompt, so English users are a first-class audience
  • Dashboard with diff and rollback gives real oversight of what an autonomous agent writes to its own long-term memory
  • Runs on a single SQLite file with no external service, and offers a live public demo for a zero-install trial

Watch out

  • The advertised eleven-client roster is mostly a bare compatibility list — only Claude Code and Antigravity have dedicated config snippets; Windsurf, Cline, OpenCode, Gemini CLI, Cherry Studio, and OpenClaw have no client-specific instructions anywhere
  • Deep-dive docs remain Chinese-only, including the full MCP tool parameter reference and the memory-audit skill files
  • No semantic search at all — search is literal substring matching, so recall quality depends entirely on the agent's own URI and keyword discipline and may degrade at scale
  • Single-maintainer project on a personal account with no visible CI or test-coverage signals — real bus-factor risk
  • The README's flagship examples are heavily companion-roleplay and NSFW-flavoured, which is a legitimate demo of the identity system but an unusual framing for teams evaluating it as agent infrastructure

Best for

  • Solo users wanting a persistent, portable, human-auditable identity layer that follows them across any MCP-capable model or client
  • Developers who want full local control on a single SQLite file with zero external services
  • Anyone who wants the agent to explicitly decide and disclose what it remembers rather than have it extracted in the background

How it integrates

Benchmark results

No sourced results yet.

Sources

Last verified 2026-07-25 · updated by discover-frameworks