garrytan/gbrain — Repo Appraisal

Overview

What it is — GBrain is a Postgres-native personal and team knowledge brain that combines hybrid RAG search, a self-wiring typed knowledge graph, and an autonomous overnight "dream cycle" for ingestion, enrichment, and consolidation.
Problem — Existing personal-knowledge tools return lists of matching pages rather than synthesized answers; agents lack a persistent, continuously-updated memory layer that can also surface what it doesn't know.
Who it's for — Individuals and teams running AI agents (OpenClaw, Hermes, Claude Code, Cursor) who need a queryable, growing institutional memory accessible over MCP or CLI.
Notable — Ships synthesis + graph traversal + gap analysis in one self-hosted package, benchmarked at P@5 49.1% / R@5 97.9% on a 240-page corpus, and reached 19K stars in 52 days as the production brain behind YC's CEO's agent deployments.

Verdict

	Rating	Summary
Quality	solid (18/24)	Exceptional documentation and engineering discipline undercut only by the absence of a tagged release and 321 open issues on a repo just 52 days old.
PAI Relevance	integrate (0.75)	TypeScript/Bun CLI with MCP stdio server drops straight into PAI's stack; synthesis, gap-analysis, and dream-cycle patterns meaningfully extend PAI's Knowledge skill with no capability PAI already owns.

Quality Assessment

18/24 — stale-risk / well-documented / high-discipline

Health: 4/8 (stale-risk)

Failed:

H1: FAIL — no tagged release; latest_release: none
H2: FAIL — no release at all, so recency criterion is moot
H6: FAIL — 321 open issues exceeds the <100 healthy-triage threshold for a repo only 52 days old
H8: FAIL — README contains no CI badge and no explicit reference to .github/workflows/; test scripts in package.json suggest local runner harnesses, not confirmed remote CI

Passed:

H3: PASS — last commit is 2026-05-27 (today), well within 6 months
H4: PASS — last commit is 2026-05-27 (today), within 30 days
H5: PASS — archived: false
H7: PASS — MIT license present

Documentation: 7/8 (well-documented)

Failed:

D8: FAIL — no explicit "Limitations," "Caveats," or "Known Issues" heading; the demo's "heads up" note describes brain-data gaps, not tool constraints

Passed:

D1: PASS — README is present and non-empty
D2: PASS — README is several thousand words, far exceeding the 1 000-byte threshold
D3: PASS — dedicated "Install" section with CLI, agent-driven, and MCP paths
D4: PASS — multiple code blocks under install and usage headings, including full CLI invocations and query/response examples
D5: PASS — references docs/INSTALL.md, docs/mcp/, and lists 30+ MCP tools by name
D6: PASS — "Search gives you raw pages. GBrain gives you the answer." appears in the opening sentence
D7: PASS — links to docs/tutorials/personal-brain.md, docs/tutorials/company-brain.md, docs/INSTALL.md, and the sibling gbrain-evals repo

Engineering Signals: 7/8 (high-discipline)

Failed:

E3: FAIL — dependency manifest section omits the dependencies field; given PGLite, embeddings, MCP, transcription, and CLI capabilities, direct dep count almost certainly exceeds 30

Passed:

E1: PASS — primary language is TypeScript
E2: PASS — package.json present with type: module, bin, and exports entries
E4: PASS — test, test:full, test:e2e, test:slow, test:heavy, and verify scripts all present; ingestion/test-harness export confirms structured test infrastructure
E5: PASS — 19,367 stars, far above the 50-star threshold
E6: PASS — ~19,367 stars over 52 days ≈ 11,200 stars/month, vastly above the 2/month floor
E7: PASS — 2,725 forks
E8: PASS — description "Garry's Opinionated OpenClaw/Hermes Agent Brain" is 45 characters and meaningful

PAI Relevance

Dimension	Score	Assessment
Harvest Value	2	The typed knowledge-graph with zero-LLM entity-edge extraction (`attended`, `works_at`, `invested_in`, `founded`, `advises`) and the gap-analysis output pattern are novel relative to PAI's file-based Knowledge skill; the dream-cycle cron architecture directly informs how PAI's overnight hooks could be composed into a consolidated enrichment loop.
Integration Readiness	2	TypeScript, Bun runtime, global CLI install via `bun install -g github:garrytan/gbrain`, MCP stdio server exposing 30+ tools — this matches PAI's stack exactly and can be added as a `gbrain serve` subprocess or `bun add`-imported library with no adapter code.
Overlap Risk	1	Partial overlap with PAI's Knowledge skill (persistent knowledge graph) and Research skill (vault queries); however, GBrain's synthesis layer, gap analysis, and autonomous dream cycle are not present in any documented PAI capability.
Gap Fill	1	PAI has a Knowledge skill but no synthesis-over-corpus layer, no gap-analysis output pattern, and no described autonomous overnight enrichment cycle; GBrain addresses these functional areas where PAI has limited or no coverage.

Composite: 0.75

What Next

Capture-to-Knowledge Pipeline (Design + manual validation stage): Deploy GBrain via docker compose up, point its ingest CLI at the existing capture sources (documents, notes, validated Haiku outputs), and schedule the dream cycle as a nightly cron. The autonomous enrichment and consolidation loop replaces the manual validation bottleneck that's currently holding the pipeline at design stage — ingested items get embedded, graph-linked, and surfaced as synthesized answers without per-item human review.
Fabric Recommender (fab v2.1): Index all Fabric patterns into GBrain with their descriptions, tags, and example invocations as knowledge nodes. Rewrite fab's pattern-matching step to query GBrain over MCP (mcp query "best pattern for <intent>") instead of doing in-context similarity at call time. The knowledge graph accumulates usage co-occurrence over time, so recommendations improve with each invocation rather than staying static.
Any active Claude Code or agent session: Add GBrain's MCP endpoint to the local MCP server config (~/.claude/mcp_servers.json) so every agent session inherits queryable project memory. The gap analysis feature — GBrain surfacing what it doesn't know — becomes a concrete signal for when to brief the agent on missing context, replacing the current pattern of re-pasting background at the start of each session.

Landscape Position

Category: Personal AI & Knowledge

In this category: tinyhumansai--openhuman (excellent, skip); related cluster entries — MemPalace--mempalace, NorthwoodsSentinel--loam, UnluckyMycologist68--palimpsest, vellum-ai--vellum-assistant, malteristo--turtleos

Standing: GBrain has the strongest documentation, the most rigorous engineering signals, and the highest adoption velocity of any personal-brain repo in the vault; it is the only one shipping benchmarked retrieval numbers, a typed knowledge graph, and a self-install agent protocol together.

Evidence Base

Density: 8/10 — Available: full README (8KB+), package.json metadata (scripts, exports, bin), star/fork/issue counts, creation and last-commit dates, license, language, archived status, landscape overlap clusters, prior appraisal scores. Missing: actual dependencies block, CI configuration file content, test file structure, any release tag or changelog.

Notes

GBrain's prior landscape entry scored it poor (4.3/24) — likely reflecting an early-stage README at the time of first ingestion. Current evidence density is 8/10 with a fully elaborated README, extensive test scripts, and 19K stars, yielding a substantially revised assessment. The 321 open issues on a 52-day-old repo is the main health risk; at this growth rate, issue triage could become a bottleneck before a first release ships. The absence of any tagged release is notable for a repo used in claimed production at scale — worth watching for a v1.0 as a quality signal.