MemoryAtlas

A persistent-memory compression system for Claude Code and other agent CLIs. Lifecycle hooks capture what the agent does during a session, an AI worker compresses those observations, and a search skill injects relevant context back into future sessions.

Deep diveClaude-Mem deep dive: what 87k stars of Claude Code memory actually costs youRead →
Storage
SQLite database stores sessions, observations, and summaries; a Chroma vector database holds embeddings for hybrid semantic + keyword (FTS5) search. Observations are addressable by ID via a local HTTP API.
Retrieval
MCP search tools with progressive disclosure: a compact `search` index (~50-100 tokens/result), a `timeline` for chronological context, and `get_observations` to fetch full details only for filtered IDs — claimed ~10x token savings by filtering before fetching.
Self-host
Self-host: moderate
License
Apache-2.0
Pricing
Open source (Apache-2.0), free to self-host — the local SQLite-backed engine has no paid component. The optional CMEM Cloud sync layer (cmem.ai) left free early access and is now priced: CMEM Cloud Pro at $199/year for solo developers, and Team Cloud at $333 per seat per month with forward-deployed-engineer onboarding. Enterprise terms are not published. · Free + paid
GitHub stars
90,288
Last release
2026-08-08
Last commit
2026-08-08
First catalogued
2026-06-28

Strengths

  • Very large community (87k+ stars, 92+ contributors) and a fast, sustained release cadence (v13.11.0, mid-July 2026)
  • Works across many agent CLIs via hooks/MCP (Claude Code, Codex, Antigravity CLI, Hermes, Copilot, OpenCode, OpenClaw gateways)
  • Progressive-disclosure retrieval keeps injected context token-light

Watch out

  • Named for Claude Code but now multi-agent across many CLIs; the supported-agent list changes often, so verify it against the README — e.g. v13.10.0 (2026-07-04) removed the Gemini CLI host integration and added Antigravity CLI (`agy`) in its place, following Google's Gemini CLI free-tier sunset (the separate Gemini LLM/observation provider is unaffected)
  • Self-reported metrics (~10x token savings, compression ratios) are vendor claims, not independently reproduced
  • Independent security audits (Feb–Mar 2026) rated the local worker API HIGH risk — unauthenticated endpoints returning cleartext API keys, plus path-traversal in MCP tools. The maintainer closed the audit issue as 'not a bug,' and a brief v12.3.3 bearer-auth fix was deliberately reverted in v12.3.5 — the worker now relies solely on localhost-only binding + CORS, which is configurable to 0.0.0.0
  • Docs site (docs.claude-mem.ai) is the primary architecture reference — verify storage/retrieval details there
  • New optional Cloud Sync (cmem.ai Pro, shipped v13.11.0–v13.12.0, July 2026) uploads observation narratives and full prompt text to a third-party (Cloudflare Workers-hosted) sync hub under your cmem.ai account — a real privacy tradeoff for teams that chose this project for its local-only storage; opt-in only
  • Cloud sync stopped being free between the 2026-07-23 and 2026-08-01 checks: what was '$0, no card required' early access is now $199/year solo, and the previously 'coming soon' Team tier landed at $333 per seat per month — a steep jump if you had built a multi-device workflow around free sync. Self-hosting remains unaffected and free

Best for

  • Developers wanting drop-in persistent session memory for Claude Code and similar coding-agent CLIs

How it integrates

Benchmark results

No sourced results yet.

Sources

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