Late at night, you ask Hermes: “So what exactly did Alice promise me last time?” It scrolls through your entire chat history and still can’t answer — because the answer was never in the conversation. It’s in last week’s meeting notes, that email you never forwarded, and a voice call. Chat memory is not world knowledge, and that’s the blind spot of every AI assistant.

Y Combinator’s President and CEO, Garry Tan, clearly agrees — he built the missing layer himself, open-sourced it, and it passed 28,000 GitHub stars within weeks. The project is called GBrain, and its production deployment runs on OpenClaw and Hermes.

GBrain in one sentence

Search gives you raw pages. GBrain gives you the answer.

GBrain is the “brain layer” Garry Tan built for his own agents — synthesis, graph traversal, and gap analysis in one box. His README lists production numbers that are hard to ignore:

• 155,795 pages of knowledge, 24,589 people, 5,340 companies
• 66 cron jobs running autonomously
• His agent ingests meetings, emails, tweets, voice calls, and raw ideas while he sleeps — enriching every person and company it encounters, fixing its own citations, and consolidating memory overnight. “I wake up smarter than when I went to bed.”

Why Hermes users should care

First, a clarification: Hermes’s built-in memory (FTS5 cross-session recall + the Honcho user model, see the hermes memory command reference) is conversational memory — great at remembering what you said and did. GBrain is world knowledge — people, companies, relationships, commitments, meeting outcomes, and the links between them.

These aren’t competitors; they’re complementary. Garry Tan’s own production setup is Hermes (or OpenClaw) + GBrain. Chat stays in chat; the brain stays in the brain.

The three killer mechanisms

1. A synthesis layer that answers, not links

Ask a traditional knowledge tool “what do I need to know before my meeting with Alice?” and you get five pages to read yourself. GBrain returns a single cited answer: “Alice runs engineering at Acme, you last spoke April 22 about pricing, three things are still open…” Every claim has a source.

Even better is its gap analysis: if the brain hasn’t seen anything about Alice in six weeks, it tells you “this may be stale, worth asking her directly” — it knows what it doesn’t know.

2. A self-wiring knowledge graph, zero LLM calls

Every page write extracts entity references and creates typed edges (works_at, invested_in, founded, advises…) with zero model calls. Ask “who works at Acme?” — a question pure vector search can’t answer — and it answers instantly.

Official benchmark (240-page Opus-generated rich-prose corpus): P@5 49.1%, R@5 97.9% — +31.4 points P@5 over its graph-disabled variant.

3. The dream cycle

GBrain runs on your schedule at night: integrating the day’s memories, repairing broken citations, deduplicating entries. You sleep, it consolidates, you wake up and just ask.

How to wire it into Hermes: three paths

Path A: Fresh install (30 minutes, recommended)

Garry’s docs are explicit that the “always on, enriching around the clock” deployment is OpenClaw or Hermes:

1. Deploy Hermes via the official Railway one-click template;
2. Install GBrain: bun install -g github:garrytan/gbrain (PGLite DB is ready in 2 seconds — no server to run);
3. Hand INSTALL_FOR_AGENTS.md to Hermes and let it finish the brain install, skill deployment, and dream-cycle setup itself.

Path B: Brain already running — connect Hermes in one command

gbrain auth create "hermes"          # mint an access token
gbrain connect https://your-brain.ngrok.app/mcp --token gbrain_xxx

gbrain connect generates the full wire-up and smoke-tests the token before handoff.

Path C: MCP direct (lightest)

GBrain ships as a standard MCP server, and Hermes has a native MCP client (stdio + HTTP, see the hermes mcp command reference):

gbrain serve --surface verbs   # local stdio — 5 memory verbs: recall / remember / entity / synthesize / forget
gbrain serve --http --port 3131  # remote mode with OAuth 2.1

Register gbrain serve as an MCP server in Hermes and you’re done — those five verbs are the agent-facing “memory interface.”

Going further: a company brain (multi-user isolation)

GBrain now supports team mode: each member gets their own slice of the brain, scoped by login, seeing only what they’re allowed to see. The author reports fuzz-testing every read path (search, list, lookup, multi-source reads) with zero leaks — matching the company-brain shape YC just put on its Request for Startups.

Pitfalls (read before installing)

1. ⚠️ Don’t install the wrong package: the npm package gbrain is an unrelated third-party package! The only supported install is bun install -g github:garrytan/gbrain (note the github: prefix). If you already hit the trap, gbrain doctor detects it and prints the fix.
2. Cost: the “always on” mode needs a real server + API spend (unlike the Codex/Claude Code path, which runs on subscriptions you already have).
3. Runtime: Bun + TypeScript — install Bun first.
4. Numbers: the production figures and benchmarks above are the author’s own README claims, not independently re-measured.

Wrap-up

If Hermes’s conversational memory already feels strong, try upgrading it to world memory: meetings, people, commitments, relationship graphs — this is the stack Garry Tan wakes up to every day. To give your Hermes the same upgrade, run bun install -g github:garrytan/gbrain and hand it the INSTALL_FOR_AGENTS.md.

Curious about Hermes’s own memory stack? See the semantic memory (LanceDB) guide.