Your AI agent is smart but forgetful. Every new session starts from zero — no memory of who you met, what you read, what you decided last Tuesday. GBrain is an open-source fix for that. Built by Garry Tan (President and CEO of Y Combinator) to power his own OpenClaw and Hermes deployments, it’s a markdown-first, Postgres-backed knowledge layer that ingests meetings, emails, tweets, and notes, then auto-wires a typed knowledge graph on top — with zero LLM calls for the graph extraction. The production brain behind Garry’s actual agents currently holds 146,646 pages, 24,585 people, 5,339 companies, and 66 autonomous cron jobs. On its own benchmark (BrainBench, a 240-page rich-prose corpus), GBrain hits P@5 49.1% and R@5 97.9%, a +31.4-point P@5 lead over the same codebase with the graph layer disabled.
This is a hands-on tutorial. You’ll install GBrain locally, import a small notes folder, run a real search, watch the knowledge graph wire itself, and connect it to Claude Code via MCP. About 20 minutes start to finish. All terminal outputs below were captured from a live install of GBrain v0.38.2.0. The repository (MIT-licensed) lives at github.com/garrytan/gbrain.
What you’re building
By the end of the tutorial, you’ll have:
- A local
~/.gbrain/brain.pglitedatabase — embedded Postgres 17 (via WASM) with pgvector, zero server config. - A small “brain repo” of markdown notes about people, companies, and concepts.
- A working hybrid-search CLI that combines vector + BM25 keyword + Reciprocal Rank Fusion (RRF), with a ZeroEntropy reranker on top by default.
- A typed knowledge graph (
works_at,founded,invested_in,attended,advises,mentions) auto-extracted from your notes. - An MCP server exposing 74 tools so Claude Code, Cursor, and Windsurf can read and write to the brain directly.
Prerequisites
- macOS or Linux (Windows users: use WSL2).
- A code editor.
- Bun ≥ 1.3.10 (the runtime GBrain ships on; the repo’s
package.jsondeclares this as the minimum engine). We’ll install it in Step 1. - An embedding API key from one of: ZeroEntropy (default), OpenAI, or Voyage. Without one, you can still install and run keyword search, but
gbrain query(hybrid + vector) will return no results. - Optional: an Anthropic API key for multi-query expansion during search.
Step 1 — Install Bun and GBrain
GBrain is written in TypeScript and runs on Bun. Install it first:
curl -fsSL https://bun.sh/install | bash
exec $SHELL # reload shell so `bun` is on PATH
bun --versionNow install GBrain. As of v0.38, the canonical install path is a single global Bun install:
bun install -g github:garrytan/gbrain
gbrain --version
# gbrain 0.38.2.0Step 2 — Initialize your brain
gbrain init --pglite provisions a local PGLite database in ~/.gbrain/. PGLite is full Postgres compiled to WASM — no server, no Docker, ready in roughly two seconds.
For this tutorial we’ll defer the embedding provider so you can follow along without an API key right away — we’ll wire it up in Step 6 when we run hybrid search:
gbrain init --pglite --no-embedding(If you’d rather configure embeddings now, set one of OPENAI_API_KEY, ZEROENTROPY_API_KEY, or VOYAGE_API_KEY in your environment before running plain gbrain init --pglite.)
Real output captured from a fresh install (truncated for brevity — there are 81 migrations from schema v1 → v85):
Setting up local brain with PGLite (no server needed)...
Schema version 1 → 85 (81 migration(s) pending)
[2] slugify_existing_pages...
[2] ✓ slugify_existing_pages
[3] unique_chunk_index...
[3] ✓ unique_chunk_index
...
Brain ready at /home/you/.gbrain/brain.pglite
0 pages. Engine: PGLite (local Postgres).You now have an empty brain. Confirm:
gbrain stats
# Pages: 0
# Chunks: 0
# Embedded: 0
# Links: 0
# Tags: 0
# Timeline: 0Step 3 — Create a tiny brain repo
The brain repo is just a directory of markdown files. Each file follows GBrain’s compiled truth + timeline pattern: a current best-understanding section on top, an append-only evidence trail below.
Important: wikilinks must use the full slug path (e.g., [[people/alice-chen]], not just [[alice-chen]]) for the graph extractor to resolve them. This is a real gotcha — I tested both forms; the short form silently produces zero links.
mkdir -p ~/my-brain/people ~/my-brain/companies ~/my-brain/concepts
cd ~/my-brainCreate a person page:
cat > people/alice-chen.md <<'EOF'
---
type: person
title: Alice Chen
tags: [founder, ai-infra]
---
Founder and CEO of [[companies/acme-ai]]. Previously staff engineer at
Google Brain. Focus area: inference optimization for small language models.
---
- 2024-03-12: Met at AI Engineer Summit. Discussed sparse MoE routing.
- 2024-09-04: Announced $12M seed led by Sequoia.
- 2025-01-18: Shipped open-source inference router on GitHub.
EOFA company page:
cat > companies/acme-ai.md <<'EOF'
---
type: company
title: Acme AI
tags: [startup, inference]
---
YC W24 inference-optimization startup. Founded by [[people/alice-chen]].
Building latency-aware routing for sub-7B models.
---
- 2024-09-04: $12M seed, led by Sequoia.
- 2025-01-18: Open-sourced their inference router.
EOFAnd a concept page:
cat > concepts/inference-optimization.md <<'EOF'
---
type: concept
title: Inference Optimization
tags: [ml-systems]
---
Techniques to reduce latency and cost when serving language models:
quantization, speculative decoding, KV-cache reuse, and request batching.
EOFStep 4 — Import the repo
gbrain import is idempotent (content-hash deduplicated). We’ll pass --no-embed so this step is deterministic for readers who don’t have an embedding key set yet — embeddings get backfilled in Step 6. Real output:
gbrain import ~/my-brain/ --no-embed[gbrain phase] import.collect_files start dir=/home/you/my-brain/ strategy=markdown
[gbrain phase] import.collect_files done 2ms files=3
Found 3 markdown files
[import.files] 3/3 (100%) imported=3 skipped=0 errors=0
Import complete (0.3s):
3 pages imported
0 pages skipped (0 unchanged, 0 errors)
3 chunks createdConfirm:
gbrain list
# companies/acme-ai company 2026-05-22 Acme AI
# concepts/inference-optimization concept 2026-05-22 Inference Optimization
# people/alice-chen person 2026-05-22 Alice ChenStep 5 — Wire the knowledge graph
For a first-time import, run the link extractor explicitly to backfill the graph from your wikilinks. This is pure regex + typed inference — zero LLM calls.
gbrain extract links --source dbReal output:
[extract.links_db] 3/3 (100%) done
Links: created 2 from 3 pages (db source)
Done: 2 links, 0 timeline entries from 3 pagesTwo typed edges were inferred from the wikilinks: alice-chen --works_at--> acme-ai (from “Founder and CEO of …”) and acme-ai --founded--> alice-chen (from “Founded by …”). The inference cascade fires in order: FOUNDED → INVESTED → ADVISES → WORKS_AT → MENTIONS. No model in the loop.
Inspect the graph directly:
gbrain graph-query people/alice-chen --depth 1
# [depth 0] people/alice-chen
# --works_at-> companies/acme-ai (depth 1)gbrain backlinks companies/acme-ai
# [
# {
# "from_slug": "people/alice-chen",
# "to_slug": "companies/acme-ai",
# "link_type": "works_at",
# "context": "Founder and CEO of [[companies/acme-ai]]...",
# "link_source": "markdown",
# ...
# }
# ]This is the difference between vector search and structured retrieval. “Who works at Acme AI?” is now a one-hop typed-edge traversal, not a similarity score. That structural channel is what drives the +31.4-point P@5 lift over the graph-disabled variant on BrainBench.
Step 6 — Run a search
GBrain ships two search verbs. gbrain search is keyword-only (BM25 on Postgres tsvector) and works without embeddings:
gbrain search "inference"
# [0.3648] companies/acme-ai -- YC W24 inference-optimization startup...
# [0.3648] people/alice-chen -- Founder and CEO of [[companies/acme-ai]]...gbrain query is the full hybrid pipeline: vector (HNSW on pgvector) + BM25 + Reciprocal Rank Fusion + optional multi-query expansion (Anthropic Haiku) + an optional ZeroEntropy reranker. It needs embeddings, which we deferred in Step 2 — wire them up now:
# Set one of: ZEROENTROPY_API_KEY (default), OPENAI_API_KEY, or VOYAGE_API_KEY
export OPENAI_API_KEY=sk-...
gbrain config set embedding_model openai:text-embedding-3-large
gbrain embed --all # one-time backfill against your embedding provider
gbrain query "who works on small-model inference?"# Set one of: ZEROENTROPY_API_KEY (default), OPENAI_API_KEY, or VOYAGE_API_KEY
export OPENAI_API_KEY=sk-...
gbrain config set embedding_model openai:text-embedding-3-large
gbrain embed --all # one-time backfill against your embedding provider
gbrain query "who works on small-model inference?"Three search modes ship out of the box — conservative, balanced, tokenmax — bundling the cost/quality knobs into one config key. Default is balanced with the ZeroEntropy reranker on. RRF formula: score = sum(1 / (60 + rank)).
Step 7 — Connect to Claude Code via MCP
The brain is more useful when an AI agent can read and write to it directly. GBrain exposes 74 tools over the Model Context Protocol via stdio. The canonical setup is one command (not a hand-edited JSON file):
claude mcp add gbrain -- gbrain serveVerify the install:
claude mcp list
# gbrain stdio gbrain serveNow ask Claude Code something like “search the brain for inference optimization” and it’ll route through the search tool and return your indexed results. The actual MCP tool names are plain snake_case: get_page, put_page, delete_page, list_pages, search, query, add_link, get_backlinks, add_tag, and 65 more.
Cursor and Windsurf use the standard MCP JSON config in their respective settings UIs. The server spec is the same:
{
"mcpServers": {
"gbrain": { "command": "gbrain", "args": ["serve"] }
}
}Claude Desktop uses claude_desktop_config.json for local stdio MCP servers with the same JSON spec. Remote HTTP MCP servers must be added through Settings → Integrations with a bearer token. See docs/mcp/CLAUDE_DESKTOP.md in the repo for the GUI walkthrough.
If you want remote access from any machine, swap stdio for HTTP:
gbrain serve --http --port 8787
# Bearer auth, default-deny CORS, two-bucket rate limit, per-request audit log.
# Postgres-only by design (PGLite is local-only).Step 8 — Let the brain run itself
GBrain ships an autopilot loop. As of v0.36.4, one command computes a dependency-ordered remediation plan, submits each step as a Minion job, re-checks the brain’s health score between steps, and refuses to spend past your cost cap:
gbrain doctor --remediate --yes --target-score 90 --max-usd 5Or run it as a daemon:
gbrain autopilot --install # cron-driven, 5-minute tickHealthy brains sleep for 60 minutes between ticks. Unhealthy ones get the full overnight cycle: sync, extract, embed, consolidate, synthesize. Three phases (synthesize, patterns, consolidate) are protected so an MCP-connected agent can’t silently burn API credits.
For ad-hoc background work, the Minions queue takes shell jobs and LLM subagent jobs side by side:
gbrain jobs submit sync --params '{}' --follow
gbrain jobs stats
gbrain jobs work --queue defaultOne PGLite caveat: gbrain jobs supervisor (the auto-restarting worker daemon) is Postgres-only. PGLite’s exclusive file lock blocks the separate worker process — the CLI rejects with a clear error if config.engine === 'pglite'. If you’re on PGLite, stick with inline --follow jobs for the tutorial, or run gbrain migrate --to supabase before standing up a persistent worker.
Routing rule: deterministic work (pull tweets, parse JSON, write a page) goes to Minions; judgment work (triage an inbox, assess priority) goes to LLM sub-agents.
What just happened, in one diagram
markdown files ──> PGLite + pgvector <──> 43 skills
(your repo, (hybrid retrieval + (HOW to use the brain;
source of truth) typed graph) RESOLVER.md routes intent)
▲ │
└────────────── agent reads/writes ──────────┘The markdown repo is the system of record. GBrain is the retrieval + graph layer over it. The agent reads and writes through both, and humans can always open any .md file and edit it directly — gbrain sync picks up the change.
Where to go next
- One-line capture (new in v0.38):
gbrain capture "the thought I want to remember"lands directly ininbox/YYYY-MM-DD-. Also accepts--file,--stdin, and webhook ingestion viagbrain serve --http /ingest. - Migrate to Supabase when your brain outgrows local (PGLite is good up to ~50K pages):
gbrain migrate --to supabase. - Ingest real data with one of the recipes: voice (Twilio + OpenAI Realtime), email + calendar, 16 embedding providers, credential gateway.
- Run the benchmarks in the sibling repo gbrain-evals: BrainBench (synthetic) and
gbrain eval longmemeval(the public LongMemEval benchmark). - Author your own skills. A skill is a fat markdown file that encodes a workflow — triggers, checks, quality gate.
gbrain check-resolvablevalidates the skill tree for reachability / MECE / DRY.
The deeper bet behind GBrain is that thin harness, fat skills beats thin skills behind a fat agent. The runtime stays small; the intelligence lives in markdown files the agent reads at decision time. Each commit you make to your brain repo is permanent context your agent inherits the next time it wakes up. The longer you run it, the smarter it gets.
Marktechpost’s Visual Explainer
Key Takeaways
- GBrain (v0.38.2.0) gives AI agents a persistent, markdown-first memory layer — built by Garry Tan to power his own OpenClaw/Hermes deployments holding 146,646 pages and 24,585 people.
- Install runs locally in ~30 minutes on PGLite (Postgres 17 compiled to WASM, zero server) and scales to Supabase or self-hosted Postgres when needed.
- Every wikilink is parsed by a regex inference cascade (
FOUNDED → INVESTED → ADVISES → WORKS_AT) that writes typed graph edges with zero LLM calls. - Hybrid search (vector + BM25 + RRF + ZeroEntropy reranker) hits P@5 49.1% / R@5 97.9% on BrainBench — a +31.4-point P@5 lift over the graph-disabled variant.
- Exposes 74 tools over MCP — wire it into Claude Code with a single
claude mcp add gbrain -- gbrain serveand your agent can read/write the brain directly.
Check out the GitHub Repo and Implementation Codes. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
Credit: Source link
