Garry's Opinionated OpenClaw/Hermes Agent Brain
Pre-1.0 at v0.32.3.0 with no formal GitHub releases, yet production-claimed at serious scale (17,888 pages, 4,383 people, 723 companies, 21 cron jobs). The repo is 37 days old with extremely rapid versioning cadence visible in the README alone (v0.25 → v0.28 → v0.32 in the excerpted text). 131 open issues is a meaningful backlog for a month-old project; it indicates active community but also rough edges. Known footguns are documented (npm name squatter, bun install -g schema migration failure in #218 and #658), which is honest engineering culture but underscores instability. Squarely alpha-to-early-beta: real enough to run in production for the author, not yet stable API surface.
Exceptional across every dimension. The README covers four distinct install paths (agent platform, standalone CLI, MCP stdio, Remote MCP OAuth 2.1), benchmarks with methodology (BrainBench P@5 49.1%, R@5 97.9% on 240-page corpus), sibling eval repo, LongMemEval integration walkthrough, embedding provider decision tree (14 recipes), and documented known bugs with issue links. Beyond the README: docs/eval-bench.md, docs/eval-capture.md, docs/integrations/embedding-providers.md, AGENTS.md, CLAUDE.md, llms.txt, llms-full.txt, and per-skill SKILL.md files. LLM-optimized doc formats are a thoughtful addition. Possibly the best-documented repo in this vault.
TypeScript with strict module exports cleanly enumerated in package.json. The scripts block reveals a serious verification suite: check:privacy, check:jsonb, check:progress, check:test-isolation, check:wasm, check:admin-build, check:admin-scope-drift, check:cli-exec, check:system-of-record, plus typecheck. Three test tiers (unit parallel, slow, e2e with Docker Compose). PGLite as the embedded Postgres engine is an interesting architectural bet — zero-infrastructure but WASM-dependent. Bun runtime is a meaningful constraint for downstream portability. Injection pattern defense is centralized (INJECTION_PATTERNS) — one source of truth pattern correctly applied. No visible CI badge in the excerpt, which is a minor gap.
Last commit is same-day as appraisal (2026-05-12). Rapid sub-version releases across a 37-day window. Known issues are documented with specific issue numbers and workarounds in the README itself, indicating the maintainer is reading and triaging user pain. The postinstall hook even prints a recovery message for the common failure mode. As active as any repo in this vault.
14,943 stars and 1,970 forks in approximately 37 days is extraordinary by any metric and reflects the author's high-profile platform (YC president). Star-to-fork ratio of ~7.6:1 suggests high consumer interest relative to contributors — consistent with a tool people want to use rather than extend. Production deployment by a prominent founder provides credibility that most personal-memory repos lack entirely.
Overall: 4.3/5
Category: Personal AI Memory
Known alternatives in vault: NorthwoodsSentinel--loam (2.2/5), UnluckyMycologist68--palimpsest (1.1/5)
Differentiation: GBrain operates at a different tier than both vault alternatives. It combines hybrid vector + BM25 search with a zero-LLM-call entity extraction layer that writes typed links (attended, works_at, invested_in, founded, advises) on every page write, forming a self-wiring knowledge graph. Backlink-boosted ranking is a retrieval primitive neither loam nor palimpsest demonstrate. Formal benchmarking (BrainBench, LongMemEval in-box) is unique in this category. MCP server with 30+ tools, OAuth 2.1 remote server, 14 embedding provider recipes, and 34 agent skills make it integration-complete where alternatives are prototypes. Alternatives offer lower adoption surface and less lock-in, but deliver far less. GBrain's main risks vs. simpler tools: Bun/PGLite dependency chain, pre-1.0 API churn, and single-maintainer bus factor.
Gap or crowd: Category has two weak entries scoring 2.2 and 1.1. GBrain fills the gap decisively and would become the dominant Personal AI Memory entry in the vault.
Score: 5/5
Harvestable: (1) Hybrid search with backlink-boosted ranking — the graph layer alone adds +31.4pp P@5 over vector-only baselines, directly applicable to any vault retrieval system. (2) Zero-LLM typed entity extraction pattern — extracting works_at, invested_in etc. without LLM calls is a cost-efficient graph-building primitive. (3) PGLite embedded Postgres pattern — zero-infrastructure relational + vector store, useful as a local-first data layer template. (4) Embedding provider abstraction (14 recipes with fallback detection via gbrain doctor) — portable to any embedding-dependent system. (5) BrainBench eval harness with session capture, NDJSON wire format, and Jaccard@k / top-1 stability metrics — extractable eval methodology for any retrieval system. (6) INJECTION_PATTERNS centralization as a prompt-injection defense primitive. (7) Overnight consolidation / cron job architecture for autonomous memory maintenance.
Integration path: Three viable paths. (A) Drop-in as the knowledge vault backend via MCP stdio — 30-minute install, 30+ tools immediately available to Claude Code, Cursor, Windsurf. (B) Remote MCP with OAuth 2.1 for multi-client access (ChatGPT, Claude Desktop, Perplexity). (C) Component extraction: the hybrid search engine, link extraction module, and embedding abstraction layer are independently importable via named exports (./search/hybrid, ./link-extraction, ./embedding).
Overlap with existing: Directly overlaps NorthwoodsSentinel--loam and UnluckyMycologist68--palimpsest, both Personal AI Memory. GBrain supersedes both in every measurable dimension; the overlapping repos add no differentiating value if GBrain is adopted.
Adoption cost: Trivial as a whole tool (documented 30-minute agent-driven install). Moderate if extracting the hybrid search or entity extraction subsystems as library components — requires unwrapping PGLite dependencies and adapting to target runtime.
GBrain is the most PAI-relevant repo seen in this vault so far. It is not a toy or a proof-of-concept — it is the production brain of a high-profile founder running 21 autonomous cron jobs across ~18K pages of personal knowledge. The architecture choices (PGLite, typed entity graph, BrainBench eval harness, agent-native install via INSTALL_FOR_AGENTS.md) are all deliberate and well-reasoned. The main risks for vault inclusion are: (1) pre-1.0 API churn — integration may require maintenance as the project evolves rapidly; (2) Bun runtime dependency — not universally available in all deployment environments; (3) single-maintainer concentration; (4) npm name squatter issue creates install friction until @garrytan/gbrain is published. None of these are disqualifying. The combination of benchmark evidence, production validation, exceptional documentation, and direct MCP integration makes this a strong acquisition for any PAI infrastructure vault. Recommend acquiring and flagging for reappraisal at v1.0 or first formal release.