What is GBrain? - Personal AI Memory Explained Simply
A plain-language guide to GBrain — Garry Tan’s open-source knowledge brain that gives the AI agent you already use durable memory with sources, corrections, and cited answers.
AI agents are great in one chat — and forgetful the next. GBrain is a way to give them memory you own, with sources you can check.
GBrain is an open-source personal knowledge brain by Garry Tan.
Tagline from the project: “Give the agent you already use a memory you control.”
It stores explicit facts with sources. You can correct them. You can withdraw them. The same memory can be shared across the agents you connect.
It is not a new ChatGPT clone. It plugs into tools you already use — Claude Code, Codex, Cursor, OpenClaw, and others.
01
The problem
Forgetful agents
02
How it works
Brain, sources, answers
03
What you’d do
Meeting prep and memory
04
How it fits
gstack, agents, LLMs
1. What problem it solves
Most AI chats:
- Forget between sessions
- Make you re-explain basics
- Return a list of notes when you wanted an answer
GBrain’s contrast (from the README): search finds pages; a brain reads them for you and writes the answer — with citations, and a heads-up about what it doesn’t know yet.
Meeting prep, simply
Imagine tomorrow’s meeting with Alice.
Typical tool: five page titles you still have to open and read.
GBrain’s pitch: a short brief — who she is, what you last talked about, what’s still open — plus sources, plus “nothing new about Alice since six weeks ago.”
That’s the product idea in one scene.
For PMs and operators
The growth angle is compounding context. Every meeting, email import, and correction can make the next ask cheaper — if the memory is explicit and owned, not trapped in one chat thread.
2. How GBrain works
The memory loop
At the simplest level:
Remember
Save a fact + source
Recall
Pull it into a new chat
Correct
Fix what was wrong
Forget
Withdraw from active use
Those verbs (remember, recall, forget, and related ones) are part of GBrain’s designed memory surface for agents.
Brain vs source (mental model)
| Word | Plain meaning |
|---|---|
| Brain | Which database / knowledge store you’re using |
| Source | A content area inside that brain (wiki, notes, even a gstack folder) |
| Page | A markdown knowledge unit the brain can retrieve |
You don’t need every detail on day one. The useful picture: one brain, organized sources, pages with provenance.
Search vs “think”
GBrain can search like a knowledge tool.
It can also synthesize: write the actual answer with citations and gap analysis (“what we don’t know yet”). That synthesis layer is optional — and it’s the part the README says changes how you use the brain.
Start keyless
You can begin with keyword memory and no extra API keys for embeddings.
Later you can add semantic search, enrichment, capture, and an overnight dream / consolidate cycle (background cleanup and connections) if you want the always-on path.
Skills and recipes
GBrain ships many skills (Markdown procedures agents follow) and recipes (integration playbooks — email, calendar, voice, and more). Philosophy: thin agent runtime, fat playbooks.
Check your understanding
What is GBrain mainly adding to an AI agent you already use?
3. What a normal person would actually do
Recommended path
Add memory to the agent you already have. Keep its identity. Don’t invent a new “personal agent repo” unless you want that optional bootstrap.
Install from GitHub (not npm)
GBrain warns clearly: the npm package named gbrain is unrelated. Install only via the project’s documented GitHub paths.
Connect your coding agent
Follow the memory-only install for Claude Code / Codex / Cursor (MCP). Verify a remember → recall → correct → forget round trip.
Use it for something real
Meeting prep. “What did we decide about pricing?” Import notes you already have. Correct bad memories when you see them.
Grow carefully
Email/calendar open loops, remote MCP access, company/team brains, overnight dream cycles — all exist, and all have trust / cost tradeoffs. Read the project’s memory-boundary docs before sharing.
Optional bigger path
Some people bootstrap a new persistent personal agent (identity files + private repo) with GBrain as memory. That’s optional — not the default recommendation.
Always-on OpenClaw / Hermes setups with continuous ingestion are powerful and expensive (compute + API usage). Start smaller unless you need that.
Honest limits
forget withdraws from active recall — it is not a promise of wiping every backup. Markdown export is not a full database backup. Shared machines and shared credentials have real trust boundaries. Cloud embeddings/synthesis send text when you configure them.
4. How it fits with gstack and LLMs
gstack (sister project)
Project framing: gstack teaches agents how to code. GBrain teaches agents everything else.
- gstack = sprint process, reviews, browser QA, ship
- GBrain = durable facts, sources, synthesis, overnight maintenance
gstack can talk to GBrain (/setup-gbrain). A gstack folder can even be a source inside a brain. They stay separate products.
Large language models
GBrain does not replace the model. The model still generates language.
GBrain is the filing cabinet + research assistant the product can query before (and while) the model writes.
Background: How does AI work? — how LLMs pick the next word, and how agent loops use tools.
You ask a question
└─ Agent (Claude Code, OpenClaw, …)
├─ Optional: recall from GBrain (facts + sources)
├─ LLM writes the reply (token by token)
└─ Optional: remember / correct for next timeQuick recap
- GBrain = memory you control for agents you already use.
- Sources + corrections beat vibes locked in one chat.
- Synthesis aims for answers, not only page lists — and admits gaps.
- Start keyless; add enrichment and always-on later if you need them.
- Pair with gstack when you want coding process + memory together.
The takeaway
GBrain is Garry Tan’s open-source knowledge brain: durable facts with provenance for the AI agent you already run. Search finds pages. A brain tries to give you the brief — and tells you what it still doesn’t know.
References
- GBrain on GitHub — source, install, docs map
- GBrain on DeepWiki — browsable code explainer
- gstack on GitHub — sister process toolkit
Related on Traction Growth
- What is Jev? — a typed decision model for routing, scoring, and software gates
- What is gstack? — virtual eng team process for AI coding agents
- How does AI work? — plain-English LLM explainer
- Home — product and growth learning from real systems
- Open source product growth case studies — deeper OSS explainers