MemoryAtlas

Local-first AI memory distributed as a Python CLI/library plus an MCP server. Stores conversation and project history as verbatim text — it explicitly does not summarize, extract, or paraphrase — and retrieves it with semantic search over a structured index where people/projects are 'wings', topics are 'rooms', and original content lives in 'drawers' so searches can be scoped rather than run flat. Bundles a temporal entity-relationship knowledge graph with validity windows.

Storage
Verbatim drawers indexed in a pluggable vector backend (ChromaDB by default; sqlite_exact for local exact-vector checks, plus opt-in Qdrant and pgvector backends behind a single storage contract). A temporal entity-relationship knowledge graph is backed by local SQLite. Nothing leaves the machine unless an external backend is explicitly configured.
Retrieval
Semantic search scoped by wing/room, with an optional hybrid pipeline (keyword boosting, temporal-proximity boosting, preference-pattern extraction) and an optional LLM-reranker over the top-20 retrieved sessions. The raw semantic path requires no API key, no cloud, and no LLM.
Self-host
Self-host: trivial
License
MIT
Pricing
Open-source (MIT) Python package, free to self-host; raw semantic retrieval needs no API key or cloud. Docs at mempalaceofficial.com; no paid tier identified. · Free / OSS
GitHub stars
58,269
Last release
2026-07-17
Last commit
2026-08-08
First catalogued
2026-06-28

Strengths

  • Verbatim, no-extraction storage — the original text is never paraphrased away, so retrieval is auditable against the source
  • Structured palace index (wings/rooms/drawers) lets searches be scoped instead of run against a flat corpus, plus a temporal SQLite knowledge graph
  • Pluggable vector backend (ChromaDB default; sqlite_exact, Qdrant, pgvector, and an opt-in Milvus backend added in v3.6.0) and a 36-tool MCP server; raw semantic recall runs with no API key or cloud
  • v3.6.0 (2026-07-17) added a secure-by-default `mempalace serve` remote/team server (TLS, bearer-token auth, read-only mode, Docker Compose/systemd templates) — a real step past single-machine-only deployment

Watch out

  • README headlines self-reported retrieval benchmarks (96.6% R@5 raw / 98.4% held-out on LongMemEval, plus LoCoMo/ConvoMem/MemBench) — record those via harvest-benchmarks with selfReported:true, not on this card
  • Very high star count (~57k) accumulated fast for a young project (created 2026-04); the README itself warns of impostor domains, so treat third-party mirrors with care
  • Hybrid/rerank quality figures are vendor-reproducible but not independently audited
  • Concurrent writers to the same palace remain restricted: v3.5.0 made the MCP server refuse a second writer rather than race the HNSW handle, and while v3.6.0 added writer-lease recovery and a 'shared daemon write-routing policy' to route writers toward one local daemon, Issue #1646's original ask (multi-agent-safe concurrent writing) is still open as of 2026-07-23

Best for

  • Local-first agent memory where verbatim, source-traceable recall and scoped semantic search matter more than fact extraction

How it integrates

Benchmark results

No sourced results yet.

Sources

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