GBrain: one memory for all your agents

1. 01

   One memory behind every agent

   Hermes, Claude Code, Codex, Cursor, ChatGPT and Grok Bot can all reach the same brain over MCP. It runs on your laptop or a server you control.

   How
2. 02

   It writes the answer, not a list

   gbrain search hands you pages. gbrain think reads them and answers with sources, plus a note on what's stale, uncited or missing.

   How
3. 03

   Built to run always on

   Garry's intended setup is Hermes on a server, running crons and ingestion around the clock. It's also the most expensive way in.

   How
4. 04

   Thin harness, fat skills

   Judgment lives in Markdown skills; anything that needs the same answer every time lives in code. A short routing file beats a giant prompt.

   How
5. 05

   Brain first, web second

   On every message the agent searches the brain before the web, answers, then writes back what it learned. Each write makes the next read better.

   How
6. 06

   Rewrite the top, append the bottom

   Each page has a summary that gets rewritten when the evidence changes and a dated, sourced timeline that is never edited. Mistakes get a correction entry.

   How
7. 07

   It works while you sleep

   A dream cycle at 2 AM creates and enriches pages, fixes missing sources and broken links, and re-indexes. Skip it and the brain slowly rots.

   How
8. 08

   Seven verbs for coding agents

   recall, remember, entity, context_pack, delta, forget and synthesize. Only synthesize is slow and costs money.

   How
9. 09

   Start keyless, opt in later

   The coding-agent setup takes about 10 minutes with no API keys. Capture, enrichment and the dream cycle are separate choices. Install from GitHub, not npm.

   How
10. 10

    It scales to a team

    One brain with several sources and a scoped login per teammate, so each person sees only what they're allowed to. Garry's recipe covers 10 to 50 people.

    How

MCP

The shared plug that lets an AI agent use outside tools. Any agent that speaks MCP (Hermes, Claude Code, Codex, Cursor, ChatGPT) can use GBrain.

Coding agent

An AI that works in your terminal, like Claude Code or Codex. Giving one a memory is Garry's recommended way to start.

LLM

Large language model: the AI model itself. GBrain keeps it away from jobs that need the same answer every time, like links and counting.

Keyless

Running with no API keys. Memory and keyword search work this way on day one; keys, capture, enrichment and the dream cycle are later opt-ins.

Bun

The tool that installs GBrain from GitHub. Install only that way, not from npm, where a package with the same name is unrelated.

Embeddings

The index that lets a hybrid query match by meaning, not just exact words. Keyword search works without them; with them, search finds synonyms and concepts.

Cron

A job that runs on a schedule. Garry's production brain runs 66 of them, from meeting sync three times a weekday to the 2 AM dream cycle.

Daemon

A program that keeps running in the background. gbrain autopilot runs the whole loop this way.

Entity

A person or company the brain keeps a page on. The dream cycle creates pages for new ones and enriches thin ones.

Slug

A page's short address in the brain, like people/alice-example. gbrain get opens a page instantly when you know it.

PGLite and Supabase

Where the pages are stored. PGLite is a small Postgres database on the same machine; Supabase is hosted Postgres, for when the brain grows or a team shares it.

Tailscale

A private network between your own devices. gbrain mcp expose puts your brain on it so ChatGPT or your phone can reach it; --funnel makes it public for cloud agents.

Hot path

Something that runs on every request and has to be quick. synthesize costs money and takes seconds to minutes, so it stays off it.

• 155,795pages
• 24,589people
• 5,340companies
• 66cron jobs

Garry's production brain, per the README.

The “heads up” is the part Garry calls the differentiator: a note on what's stale, uncited, contradictory or missing. gbrain search gives you pages; gbrain think gives you the answer and the gaps.

1. 01

   You

   Message Hermes from anywhere. Or don't: it keeps working on its schedule.

2. 02

   Hermes, on a server

   Always on. Runs the crons, collectors and the 2 AM dream cycle.

3. 03

   gbrain serve

   Started by Hermes as a local subprocess. No separate server, tunnel or token.

4. 04

   The brain

   PGLite on the same box, or Supabase when it grows. Your pages, your keys.

Setting it up

1. 01

   Deploy Hermes

   One click on Railway. Hermes is Nous Research's open-source agent that runs on a server, so it keeps working with your laptop shut.

2. 02

   Hand it the install doc

   Paste one instruction into Hermes pointing at GBrain's INSTALL_FOR_AGENTS.md. Hermes installs the brain itself.

3. 03

   It starts keyless

   Memory works on day one with no API keys. The install doesn't switch anything else on.

4. 04

   Opt in, one at a time

   API keys, automatic capture, paid enrichment and the nightly dream cycle are separate choices you make later.

The trade-off: it's the most expensive way in, a server with 8 GB+ of memory plus model usage that grows with how hard the agent works. Garry's advice: start with a coding agent to feel it first, or start here if you want the full experience from day one.

1. Fat skills

   Markdown procedures that hold judgment, process and domain knowledge. About 90% of the value.

2. Thin harness

   About 200 lines. JSON in, text out. Read-only by default.

3. Your app

   The deterministic floor: query the database, read a doc, search, timeline.

1. 01

   Signal

   A message, meeting, email, tweet or link arrives.

2. 02

   Search

   Brain first, before any web or paid API.

3. 03

   Respond

   Answer with that context.

4. 04

   Write

   Rewrite the page's summary, append a dated timeline entry.

5. 05

   Auto-link

   Typed links and backlinks from page references. No LLM calls.

6. 06

   Sync

   Index the change so the next search finds it.

next signal arrives, and the agent is a little smarter

Brain-first lookup, in this order

1. 01Keyword searchgbrain searchFast. Works on day one with no embeddings.
2. 02Hybrid querygbrain queryKeyword + meaning. Finds synonyms and concepts.
3. 03Direct pagegbrain getInstant when you know the slug, like people/alice-example.
4. 04The webfallback onlyOnly if the brain had nothing. External APIs fill gaps; they don't start from scratch.

• Compiled truth: rewrite the whole section with the new evidence folded in. Never stack old and new.
• Timeline: append a dated entry with its source. Never edit old entries.
• Wrong fact: add a “Correction: …” entry, then rewrite the summary.

Why it works: six months of entries compress into a 30-second read, and search ranks the summary above the timeline, so the freshest picture surfaces first. Don't skip the Assessment section: your own read on a person is the one thing no API can give you.

The dream cycle, nightly at 2 AM

1. 01

   Entity sweep

   Read today's conversations. New person or company? Create a page. Thin page? Enrich it. Known? Add today's mentions.

2. 02

   Fix citations

   Find timeline entries with no source and fix them. Repair broken links.

3. 03

   Consolidate

   Patterns across conversations become durable knowledge. Summaries get rewritten from new evidence.

4. 04

   Sync and embed

   Re-index everything so tomorrow's searches see it.

It ships as one command, gbrain dream, and gbrain autopilot runs the whole loop as a background daemon. Skip the dream and the brain slowly rots: broken links, facts without sources, stale summaries. It's also where the always-on cost lives, so it's opt-in.

• recallread · fast

  Get saved facts about a topic or entity.

• rememberwrite · fast

  Save one fact, with where it came from.

• entityread · fast

  Everything on one person or company. No LLM, under 100 ms.

• context_packread · fast

  A bundle of context for up to 8 entities at the start of a session.

• deltaread · fast

  What changed since you last checked.

• forgetwrite · fast

  Withdraw a fact from active memory. History and backups may remain.

• synthesizeread · $$ slow

  A written, cited answer. Costs money and takes seconds to minutes. Never on a hot path.

bun install -g github:garrytan/gbrain
gbrain init --pglite --no-embedding
claude mcp add gbrain -- gbrain serve --surface verbs
# Codex: codex mcp add gbrain -- gbrain serve --surface verbs

1. Memory for Claude Code or Codex: About 10 minutes. Free to start: keyless, no API keys. Needs: A Mac or Linux terminal and Bun.
2. Use it from anywhere: About 15 minutes after the above. Free (Tailscale), plus whatever models you add. Needs: The computer stays on.
3. Memory for Grok Bot: About 20 minutes. Free to start: keyless. Needs: GBrain on a computer that stays on. No such computer? Install it inside the Bot's /workspace instead..
4. Hermes, always on: About 30 minutes (Hermes does the work), plus the deploy. Highest: a server (8 GB+ RAM) plus API usage that scales with how hard it works. Needs: Hermes on a server (one-click Railway template).
5. A company brain for a team: About 90 minutes on top of the personal setup. Under $100/month of AI for 25 people, plus hosting. Needs: Postgres (Supabase), an HTTPS server, one login per teammate.
6. Just search my notes: 5 minutes. Free. Needs: A terminal and Bun.

The botmaster pattern: onboard each teammate yourself

1. 01

   Pre-populate their slice

   Before they log in: a one-page profile, 5–10 frameworks that are theirs, links they care about, 2–3 example pages. About 20 minutes.

2. 02

   Walk them through 2–3 wow flows

   A synthesis answer across sources. A question it can't answer, so they see it admit the gap. A meeting they tell it about, filed and linked.

3. 03

   Then hand over the DM

   Only after the wow lands. Give access first and most people send one generic query and never come back.

About 45 minutes per person. Garry: compared to the cost of an internal tool nobody adopts, it's the best 45 minutes you'll spend. Expected bill for 25 people: about $40/month embeddings and $50/month model calls, plus hosting.

1. 01

   Brain first, web second

   Search your own brain before any web or paid lookup. A LinkedIn scrape gives a job title; the brain gives the relationship.

2. 02

   Rewrite the summary, append the timeline

   The top of a page is always the current picture. The timeline below is never edited, only added to.

3. 03

   Correct with a new entry

   Wrong fact? Don't edit history. Add “Correction: …” and rewrite the summary.

4. 04

   Every claim traces to a source

   Each line in the summary should point back to a dated, sourced timeline entry.

5. 05

   Write your own assessment

   “Sharp leader, watch for burnout” is what no API can give you. It's what makes the page better than LinkedIn.

6. 06

   Dream every night

   Skip the nightly cycle and the brain slowly rots: broken links, sourceless facts, stale summaries.

7. 07

   Respect quiet hours

   Nothing pings you from 11 PM to 8 AM. Held items fold into the morning briefing, even when you travel.

8. 08

   Twice means skill + cron

   Do it by hand for 3–10 items, fix the quality, write it as a skill, then schedule it.

9. 09

   One owner per page type

   Two skills creating the same kind of page is a bug. That's how you get three pages for one person.

10. 10

    Deterministic links

    Build Slack or email links from API data. LLMs hallucinate URLs.

11. 11

    Keep CLAUDE.md to pointers

    Garry's grew to 20,000 lines and the model's attention degraded. About 200 lines of pointers fixed it.

12. 12

    Test recall in a new conversation

    Remembering inside the same chat proves nothing. Start fresh and watch it actually call the brain.

Related on this site: Company Brain, the Slack-only brain compared above. Compound Engineering, where the agent writes down what it learned so the next job starts smarter. autoresearch, another setup that keeps working while you sleep.

• garrytan/gbrain (MIT) by Garry Tan: README and docs as of Sept 26, 2026, including the compiled-truth pattern, brain-first lookup, operational disciplines, reference cron schedule, memory verbs spec, search modes and the company-brain tutorial.
• Thin Harness, Fat Skills, Garry's essay (YC Spring 2026 talk).
• gbrain-evals for benchmark receipts · Garry Tan on X
• The Dana Ruiz page demo, the setup picker and the comparison table are FakeNerd's, built from the docs.