A markdown-repo-native memory layer: a git repo of frontmatter-tagged markdown pages is the canonical store, synced into either embedded PGLite or Postgres/pgvector as a disposable, rebuildable index. A contract-first BrainEngine interface (~47 operations) is implemented identically by both storage engines, so the CLI and MCP server are generated from one source. On top of storage sits an optional synthesis layer (`gbrain think`) that composes cited answers with explicit gap analysis, plus a cron-driven 'dream cycle' that enriches, dedupes, and detects contradictions overnight.
- Storage
- Markdown files with YAML frontmatter in a git repo are the system of record; the database is a derived, rebuildable cache — either embedded PGLite (zero-config, no Docker, recommended to ~50K pages) or Postgres/pgvector via Supabase or self-host for shared deployments. Deletes in git become soft-deletes in the DB, and `gbrain rebuild` reconstructs the entire index from the repo.
- Retrieval
- Hybrid search layering HNSW vector similarity (pgvector), BM25 keyword, and reciprocal-rank fusion, plus a zero-LLM-call knowledge graph of typed edges (works_at, invested_in, founded, advises) extracted from wikilinks at write time for multi-hop traversal, with an optional ZeroEntropy cross-encoder reranker.
- Self-host
- Self-host: trivial
- License
- MIT
- Pricing
- Free and open source, self-hosted only — no GBrain-branded managed tier exists. The company-brain and personal-brain paths run on third-party infra you pay for directly (Render, Railway, Fly.io, Supabase), not a GBrain SaaS. · Free / OSS
- GitHub stars
- 27,198
- Last release
- 2026-07-24
- Last commit
- 2026-07-25
- First catalogued
- 2026-07-25
Strengths
- Independently adoptable as coding-agent memory: the featured starting point is a two-command local MCP setup with zero OpenClaw or Hermes dependency
- Git-native, human-readable, disaster-recoverable storage — markdown is the system of record and the database is an explicitly rebuildable cache
- Retrieval genuinely layers vector, BM25, a zero-LLM-cost knowledge graph, and a reranker, with a documented architectural rationale for each
- Synthesis layer returns cited prose plus an explicit statement of what the brain does not know, rather than a list of matching chunks
- Unusually transparent benchmark culture: the sibling gbrain-evals repo publishes methodology, sealed answer keys, and unflattering results alongside favourable ones
Watch out
- Every benchmark number (BrainBench P@5 49.1% / R@5 97.9%, LongMemEval 97.6% recall@5) is authored, run, and published by the same maintainer on an LLM-generated fictional corpus — treat as vendor claims, not independent results
- Positioned as much as a full autonomous daemon (66 cron jobs, 43 skills, a dream cycle, a sub-agent job queue) as a memory layer; the MCP and CLI surfaces are genuinely standalone, but confirm in practice that adopting it doesn't pull in the daemon
- Single-maintainer governance: the GitHub description reads 'Garry's Opinionated…', the LICENSE copyright is held personally, and several credited contributors appear to be automated agent accounts
- No registry-distributed package — install is a live git checkout via `bun install -g github:garrytan/gbrain`, and the `gbrain` name on npm belongs to an unrelated GPU-ML library, so there is no versioned artifact to pin
- PGLite, the 2-second zero-config default, is documented as crashing on macOS 26.x (Tahoe) on Apple Silicon, requiring a fallback to native Homebrew Postgres + pgvector
Best for
- Developers on Claude Code, Codex, or Cursor who want git-backed markdown memory wired in via MCP in two commands
- Power users building a personal brain over meetings, email, and notes who want graph-aware recall and synthesized cited answers
- Small teams wanting a self-hosted, OAuth-scoped shared institutional memory
How it integrates
Benchmark results
No sourced results yet.
Sources
- Repo metadata — 27,100 stars, MIT, default branch master, pushed 2026-07-25; description reads 'Garry's Opinionated OpenClaw/Hermes Agent Brain' (vendor)
- README — install paths (agent-installed / CLI standalone / MCP), capabilities, BrainBench summary, license (vendor)
- Confirms MCP works standalone in Claude Code via stdio with no OpenClaw or Hermes required — the inclusion-critical fact (vendor)
- Storage model — markdown git repo is system of record, database is a derived rebuildable cache (vendor)
- Retrieval architecture (hybrid vector + BM25 + RRF + graph + reranker) and self-reported BrainBench comparison table (vendor)
- BrainBench/LongMemEval test suite — confirms benchmark numbers are self-published by the same maintainer (vendor)
- Confirms the 'gbrain' npm name is an unrelated GPU-ML library — no registry-distributed release for this project (third-party)
Last verified 2026-07-25 · updated by discover-frameworks