GBrain

Open source personal knowledge brain. Postgres + pgvector + hybrid search that actually works.

# You have 342 markdown files scattered across repos. GBrain makes them searchable.
gbrain import ~/git/brain/
# Imported 342 files (1,847 chunks) into Supabase. Embedding...

gbrain query "what are our biggest risks right now?"

strategy/competitive-moats (concept) score=0.0312
  A durable competitive advantage comes from compounding effects that
  are hard to replicate. Network effects, switching costs, scale...

meetings/2025-03-board-prep (source) score=0.0298
  Board discussion covered market positioning against three emerging
  competitors. Key concern: pricing pressure in enterprise segment...

people/jane-chen (person) score=0.0251
  VP Strategy. Led the competitive analysis project in Q1. Published
  internal framework for evaluating competitive threats...

Hybrid search finds knowledge by meaning, not just keywords. "Biggest risks" matches pages about competitive moats, board prep, and strategy leads even when the exact phrase doesn't appear. That's the point.

Why this exists

You have a brain full of knowledge. It lives in markdown files, meeting notes, CRM exports, Obsidian vaults, Notion databases. It's scattered, unsearchable, and going stale.

Search is the bottleneck. Keyword search misses semantic matches. Vector search misses exact names and phrases. Neither connects related ideas across documents.

GBrain fixes this with hybrid search that combines both approaches, plus a knowledge model that treats every page like an intelligence assessment: compiled truth on top (your current best understanding, rewritten when evidence changes), append-only timeline on the bottom (the evidence trail that never gets edited).

AI agents maintain the brain. You ingest a document and the agent updates every entity mentioned, creates cross-reference links, and appends timeline entries. MCP clients query it. The intelligence lives in fat markdown skills, not application code.

Try it: your files, searchable in 90 seconds

GBrain doesn't ship with demo data. It finds YOUR markdown and makes it searchable.

Act 1: Discovery. GBrain scans your machine for markdown repos.

=== GBrain Environment Discovery ===

  ~/git/brain (2.3GB, 342 .md files, 87 binary files)
    Type: Plain markdown (ready for import)

  ~/Documents/obsidian-vault (180MB, 1,203 .md files, 0 binary files)
    Type: Obsidian vault (wikilink conversion available)

=== Discovery Complete ===

Act 2: Import. Your files move from the repo into Supabase.

gbrain import ~/git/brain/
# Imported 342 files into Supabase (1,847 chunks). Embedding in background...

gbrain stats
# Pages: 342, Chunks: 1,847, Embedded: 0 (embedding...), Links: 0

Act 3: Search. The agent picks a query from your actual content.

# The agent reads your corpus and picks a relevant query
gbrain query "what do we know about competitive dynamics?"
# 3 results, scored by hybrid search (vector + keyword + RRF fusion)

# 30 seconds later, embeddings finish:
gbrain stats
# Pages: 342, Chunks: 1,847, Embedded: 1,847, Links: 0

# Now semantic search is live too
gbrain query "what are our biggest risks right now?"
# Finds pages about moats, board prep, and strategy -- by meaning, not keywords

Your file count will be different. Your queries will be different. The agent picks them based on what it imported. That's the point: this is YOUR brain, not a demo.

Install

Prerequisites

GBrain needs three things to run:

Dependency	What it's for	How to get it
Supabase account	Postgres + pgvector database	supabase.com (Pro tier, $25/mo for 8GB)
OpenAI API key	Embeddings (text-embedding-3-large)	platform.openai.com/api-keys
Anthropic API key	Multi-query expansion + LLM chunking (Haiku)	console.anthropic.com

Set the API keys as environment variables:

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...

The Supabase connection URL is configured during gbrain init. The OpenAI and Anthropic SDKs read their keys from the environment automatically.

Without an OpenAI key, search still works (keyword only, no vector search). Without an Anthropic key, search still works (no multi-query expansion, no LLM chunking).

With OpenClaw (recommended)

If you're running OpenClaw, paste this to set up your brain. Make sure your API keys are set in the environment first.

You: "Set up gbrain (https://github.com/garrytan/gbrain) as my knowledge brain.
      I need you to:
      1. Make sure bun is installed (curl -fsSL https://bun.sh/install | bash), then run: bun add gbrain
      2. Run: gbrain init --supabase (follow the wizard to connect my Supabase database)
      3. Scan ~/git/ and ~/Documents/ for markdown repos, pick the best one, and run: gbrain import <path>
      4. Run a query against the imported data to prove search works -- pick the query based on what you imported
      5. Read https://github.com/garrytan/gbrain/blob/master/docs/GBRAIN_RECOMMENDED_SCHEMA.md and offer to restructure my knowledge base"

OpenClaw will install the package, walk through the Supabase connection wizard, discover your markdown files, import them into Supabase, prove search works with a query from your data, and learn the 7 brain skills (ingest, query, maintain, enrich, briefing, migrate, install).

After setup, you talk to your brain through OpenClaw:

You: "Search the brain for everything we know about [topic from your data]"
You: "Ingest my meeting notes from today"
You: "Give me a briefing for my meetings tomorrow"
You: "How many pages are in the brain now?"

OpenClaw reads the skill files in skills/, figures out which gbrain commands to run, and does the work. You never touch the CLI directly unless you want to.

GBrain keeps your brain current automatically. After setup, gbrain sync --watch polls your git repo and imports only what changed. Binary files (images, PDFs, audio) can be moved to Supabase Storage with gbrain files sync to slim down your git repo.

With ClawHub

clawhub install gbrain

ClawHub installs the bundle plugin, configures the MCP server, and auto-runs the setup skill. Each brain should have its own Supabase project (one project per person or team).

Standalone CLI

bun add -g gbrain

As a library

bun add gbrain

import { PostgresEngine } from 'gbrain';

All paths require a Postgres database with pgvector. Supabase Pro ($25/mo) is the recommended zero-ops option.

Upgrade

Upgrade depends on how you installed:

# Installed via bun (standalone or library)
bun update gbrain

# Installed via ClawHub
clawhub update gbrain

# Compiled binary
# Download the latest from https://github.com/garrytan/gbrain/releases

After upgrading, run gbrain init again to apply any schema migrations (idempotent, safe to re-run).

Setup

After installing via CLI or library path, run the setup wizard:

# Guided wizard: auto-provisions Supabase or accepts a connection URL
gbrain init --supabase

# Or connect to any Postgres with pgvector
gbrain init --url postgresql://user:pass@host:5432/dbname

The init wizard:

1. Checks for Supabase CLI, offers auto-provisioning
2. Falls back to manual connection URL if CLI isn't available
3. Runs the full schema migration (tables, indexes, triggers, extensions)
4. Verifies the connection and confirms the database is ready for import

Config is saved to ~/.gbrain/config.json with 0600 permissions.

OpenClaw users skip this step. The orchestrator runs the wizard for you during install.

First import

# Import your markdown wiki (auto-chunks and auto-embeds)
gbrain import /path/to/brain/

# Skip embedding if you want to import fast and embed later
gbrain import /path/to/brain/ --no-embed

# Backfill embeddings for pages that don't have them
gbrain embed --stale

Import is idempotent. Re-running it skips unchanged files (compared by SHA-256 content hash). Progress bar shows status. ~30s for text import of 7,000 files, ~10-15 min for embedding.

The knowledge model

Every page in the brain follows the compiled truth + timeline pattern:

---
type: concept
title: Do Things That Don't Scale
tags: [startups, growth, pg-essay]
---

Paul Graham's argument that startups should do unscalable things early on.
The most common: recruiting users manually, one at a time. Airbnb went
door to door in New York photographing apartments. Stripe manually
installed their payment integration for early users.

The key insight: the unscalable effort teaches you what users actually
want, which you can't learn any other way.

---

- 2013-07-01: Published on paulgraham.com
- 2024-11-15: Referenced in batch W25 kickoff talk
- 2025-02-20: Cited in discussion about AI agent onboarding strategies

Above the --- separator: compiled truth. Your current best understanding. Gets rewritten when new evidence changes the picture. Below: timeline. Append-only evidence trail. Never edited, only added to.

The compiled truth is the answer. The timeline is the proof.

How search works

Query: "when should you ignore conventional wisdom?"
         |
    Multi-query expansion (Claude Haiku)
    "contrarian thinking startups", "going against the crowd"
         |
    +----+----+
    |         |
  Vector    Keyword
  (HNSW     (tsvector +
  cosine)    ts_rank)
    |         |
    +----+----+
         |
    RRF Fusion: score = sum(1/(60 + rank))
         |
    4-Layer Dedup
    1. Best chunk per page
    2. Cosine similarity > 0.85
    3. Type diversity (60% cap)
    4. Per-page chunk cap
         |
    Stale alerts (compiled truth older than latest timeline)
         |
    Results

Keyword search alone misses conceptual matches. "Ignore conventional wisdom" won't find an essay titled "The Bus Ticket Theory of Genius" even though it's exactly about that. Vector search alone misses exact phrases when the embedding is diluted by surrounding text. RRF fusion gets both right. Multi-query expansion catches phrasings you didn't think of.

Database schema

10 tables in Postgres + pgvector:

pages                    The core content table
  slug (UNIQUE)          e.g. "concepts/do-things-that-dont-scale"
  type                   person, company, deal, yc, civic, project, concept, source, media
  title, compiled_truth, timeline
  frontmatter (JSONB)    Arbitrary metadata
  search_vector          Trigger-based tsvector (title + compiled_truth + timeline + timeline_entries)
  content_hash           SHA-256 for import idempotency

content_chunks           Chunked content with embeddings
  page_id (FK)           Links to pages
  chunk_text             The chunk content
  chunk_source           'compiled_truth' or 'timeline'
  embedding (vector)     1536-dim from text-embedding-3-large
  HNSW index             Cosine similarity search

links                    Cross-references between pages
  from_page_id, to_page_id
  link_type              knows, invested_in, works_at, founded, references, etc.

tags                     page_id + tag (many-to-many)

timeline_entries         Structured timeline events
  page_id, date, source, summary, detail (markdown)

page_versions            Snapshot history for compiled_truth
  compiled_truth, frontmatter, snapshot_at

raw_data                 Sidecar JSON from external APIs
  page_id, source, data (JSONB)

files                    Binary attachments in Supabase Storage
  page_slug (FK)         Links to pages (ON UPDATE CASCADE)
  storage_path, content_hash, mime_type, metadata (JSONB)

ingest_log               Audit trail of import/ingest operations

config                   Brain-level settings (embedding model, chunk strategy, sync state)

Indexes: B-tree on slug/type, GIN on frontmatter/search_vector, HNSW on embeddings, pg_trgm on title for fuzzy slug resolution.

Chunking

Three strategies, dispatched by content type:

Recursive (timeline, bulk import): 5-level delimiter hierarchy (paragraphs, lines, sentences, clauses, words). 300-word chunks with 50-word sentence-aware overlap. Fast, predictable, lossless.

Semantic (compiled truth): Embeds each sentence, computes adjacent cosine similarities, applies Savitzky-Golay smoothing to find topic boundaries. Falls back to recursive on failure. Best quality for intelligence assessments.

LLM-guided (high-value content, on request): Pre-splits into 128-word candidates, asks Claude Haiku to identify topic shifts in sliding windows. 3 retries per window. Most expensive, best results.

Commands

SETUP
  gbrain init [--supabase|--url <conn>]     Create brain (guided wizard)
  gbrain upgrade                            Self-update

PAGES
  gbrain get <slug>                         Read a page (supports fuzzy slug matching)
  gbrain put <slug> [< file.md]             Write/update a page (auto-versions)
  gbrain delete <slug>                      Delete a page
  gbrain list [--type T] [--tag T] [-n N]   List pages with filters

SEARCH
  gbrain search <query>                     Keyword search (tsvector)
  gbrain query <question>                   Hybrid search (vector + keyword + RRF + expansion)

IMPORT/EXPORT
  gbrain import <dir> [--no-embed]          Import markdown directory (idempotent)
  gbrain sync [--repo <path>] [flags]       Git-to-brain incremental sync
  gbrain export [--dir ./out/]              Export to markdown (round-trip)

FILES
  gbrain files list [slug]                  List stored files
  gbrain files upload <file> --page <slug>  Upload file to storage
  gbrain files sync <dir>                   Bulk upload directory
  gbrain files verify                       Verify all uploads

EMBEDDINGS
  gbrain embed [<slug>|--all|--stale]       Generate/refresh embeddings

LINKS + GRAPH
  gbrain link <from> <to> [--type T]        Create typed link
  gbrain unlink <from> <to>                 Remove link
  gbrain backlinks <slug>                   Incoming links
  gbrain graph <slug> [--depth N]           Traverse link graph (recursive CTE, default depth 5)

TAGS
  gbrain tags <slug>                        List tags
  gbrain tag <slug> <tag>                   Add tag
  gbrain untag <slug> <tag>                 Remove tag

TIMELINE
  gbrain timeline [<slug>]                  View timeline entries
  gbrain timeline-add <slug> <date> <text>  Add timeline entry

ADMIN
  gbrain stats                              Brain statistics
  gbrain health                             Health dashboard (embed coverage, stale, orphans)
  gbrain history <slug>                     Page version history
  gbrain revert <slug> <version-id>         Revert to previous version
  gbrain config [get|set] <key> [value]     Brain config
  gbrain serve                              MCP server (stdio)
  gbrain call <tool> '<json>'               Raw tool invocation
  gbrain --tools-json                       Tool discovery (JSON)

Using as a library

GBrain is library-first. The CLI and MCP server are thin wrappers over the engine.

import { PostgresEngine } from 'gbrain';

const engine = new PostgresEngine();
await engine.connect({ database_url: process.env.DATABASE_URL });
await engine.initSchema();

// Write a page
await engine.putPage('concepts/superlinear-returns', {
  type: 'concept',
  title: 'Superlinear Returns',
  compiled_truth: 'Paul Graham argues that returns in many fields are superlinear...',
  timeline: '- 2023-10-01: Published on paulgraham.com',
});

// Hybrid search
const results = await engine.searchKeyword('startup growth');

// Typed links
await engine.addLink('concepts/superlinear-returns', 'concepts/do-things-that-dont-scale', '', 'references');

// Graph traversal
const graph = await engine.traverseGraph('concepts/superlinear-returns', 3);

// Health check
const health = await engine.getHealth();
// { page_count: 10, embed_coverage: 1.0, stale_pages: 0, orphan_pages: 10 }

The BrainEngine interface is pluggable. See docs/ENGINES.md for how to add backends.

MCP server

Add to your Claude Code or Cursor MCP config:

{
  "mcpServers": {
    "gbrain": {
      "command": "gbrain",
      "args": ["serve"]
    }
  }
}

30 tools generated from the contract-first operations.ts: page CRUD, search, tags, links, timeline, admin, sync, raw data, file management, and more.

Every tool is generated from the same operation definitions as the CLI. Parity tests verify structural identity.

Skills

Fat markdown files that tell AI agents HOW to use gbrain. No skill logic in the binary.

Skill	What it does
ingest	Ingest meetings, docs, articles. Updates compiled truth (rewrite, not append), appends timeline, creates cross-reference links across all mentioned entities.
query	3-layer search (keyword + vector + structured) with synthesis and citations. Says "the brain doesn't have info on X" rather than hallucinating.
maintain	Periodic health: find contradictions, stale compiled truth, orphan pages, dead links, tag inconsistency, missing embeddings, overdue threads.
enrich	Enrich pages from external APIs. Raw data stored separately, distilled highlights go to compiled truth.
briefing	Daily briefing: today's meetings with participant context, active deals with deadlines, time-sensitive threads, recent changes.
migrate	Universal migration from Obsidian (wikilinks to gbrain links), Notion (stripped UUIDs), Logseq (block refs), plain markdown, CSV, JSON, Roam.
setup	Set up GBrain from scratch: auto-provision Supabase via CLI, AGENTS.md injection, import, sync. Target TTHW < 2 min.

Architecture

CLI / MCP Server
     (thin wrappers, identical operations)
              |
      BrainEngine interface
       (pluggable backend)
              |
     +--------+--------+
     |                  |
PostgresEngine     SQLiteEngine
  (ships v0)       (designed, community PRs welcome)
     |
Supabase Pro ($25/mo)
  Postgres + pgvector + pg_trgm
  connection pooling via Supavisor

Embedding, chunking, and search fusion are engine-agnostic. Only raw keyword search (searchKeyword) and raw vector search (searchVector) are engine-specific. RRF fusion, multi-query expansion, and 4-layer dedup run above the engine on SearchResult[] arrays.

Storage estimates

For a brain with ~7,500 pages:

Component	Size
Page text (compiled_truth + timeline)	~150MB
JSONB frontmatter + indexes	~70MB
Content chunks (~22K, text)	~80MB
Embeddings (22K x 1536 floats)	~134MB
HNSW index overhead	~270MB
Links, tags, timeline, versions	~50MB
Total	~750MB

Supabase free tier (500MB) won't fit a large brain. Supabase Pro ($25/mo, 8GB) is the starting point.

Initial embedding cost: ~$4-5 for 7,500 pages via OpenAI text-embedding-3-large.

Docs

• GBRAIN_RECOMMENDED_SCHEMA.md -- The recommended brain schema: MECE directories, compiled truth + timeline, enrichment pipelines, resolver decision tree
• GBRAIN_V0.md -- Full product spec, all architecture decisions, every option considered
• ENGINES.md -- Pluggable engine interface, capability matrix, how to add backends
• SQLITE_ENGINE.md -- Complete SQLite engine plan with schema, FTS5, vector search options

Contributing

See CONTRIBUTING.md. Run bun test for unit tests. For E2E tests against real Postgres+pgvector: docker compose -f docker-compose.test.yml up -d then DATABASE_URL=postgresql://postgres:postgres@localhost:5434/gbrain_test bun run test:e2e.

Welcome PRs for:

• SQLite engine implementation
• Additional migration sources (Logseq, Roam, Notion)
• New enrichment API integrations
• Performance optimizations

License

MIT