MemoryAtlas

A persistent-memory server for AI coding agents, built on the `iii` engine and extending Karpathy's LLM-wiki pattern with confidence scoring, lifecycle, knowledge graphs, and hybrid search. Exposes 53 MCP tools and 12 auto-hooks; zero external databases required.

Storage
Embedded SQLite plus the iii-engine; optional knowledge-graph extraction (entities + BFS traversal). 4-tier memory consolidation with versioning, supersession, TTL expiry, contradiction detection, and importance-based eviction. Secrets and `<private>` tags are stripped before storage.
Retrieval
Hybrid BM25 + vector + graph search fused with Reciprocal Rank Fusion (RRF, k=60), session-diversified (max 3 per session). SessionStart hook loads a project profile (top concepts, files, patterns) within a token budget (default 2000) and injects it into the conversation.
Self-host
Self-host: trivial
License
Apache-2.0
Pricing
Open source (Apache-2.0), free to self-host; npm package @agentmemory/agentmemory; no managed tier advertised · Free / OSS
GitHub stars
24,202
Last release
2026-06-07
Last commit
2026-06-28
First catalogued
2026-06-28

Strengths

  • Zero external databases (embedded SQLite) lowers ops burden
  • Hybrid BM25 + vector + graph retrieval with RRF fusion; multi-script tokenization (Greek, Cyrillic, Hebrew, Arabic, CJK with optional segmenters)
  • Active release cadence and broad agent support via MCP/hooks (Claude Code, Codex, Copilot CLI, Cursor, Gemini CLI, Hermes, OpenCode, Cline)
  • Memory governance: audit trail, git snapshots/diff, team namespacing, citation provenance

Watch out

  • Self-reported benchmark numbers (95.2% retrieval R@5, 92% fewer tokens) are vendor claims, not independently reproduced — treat the README's "#1" framing as marketing
  • Built on the `iii` engine (github.com/iii-hq/iii) — reviewer should confirm iii's maturity and license independently
  • Native Windows engine setup is manual; `agentmemory connect` unsupported on Windows (WSL2 recommended)

Best for

  • Coding agents needing a self-contained, no-external-DB persistent memory with hybrid search and team/namespacing

Benchmark results

BenchmarkValueBackboneTrustSource
longmemeval95.2 recallSelf-reportedrohitg00 (agentmemory authors)

Sources

Last verified 2026-06-28 · updated by refresh-framework-cards