gbrain: the company brain in Markdown + git, popularized by Garry Tan
A plain-English guide to gbrain — the LLM-wiki variant organized around how a company actually thinks: people, projects, decisions. Popularized by Garry Tan (YC). Your team gets one Markdown-native brain that every agent can read.
Short version: gbrain is the company-brain flavor of the Karpathy LLM-wiki pattern, organized around people, projects, and decisions — the shape of how a company actually thinks. Your agent maintains the wiki in Markdown; every employee and every AI tool reads the same brain. Popularized by Garry Tan (CEO of YC); inspired by Farza at F Inc's earlier implementation off Karpathy's sketch.
What is gbrain?
gbrain gives your company a single living brain in Markdown, maintained by AI agents, stored in git. Unlike general LLM-wiki implementations, gbrain's structure is opinionated for a business: People pages (your team, customers, investors), Project pages (active initiatives, deliverables, owners), Decision pages (what you decided, why, when, alternatives considered), and System pages (how the company works).
When Garry Tan started showing off how YC runs on a git-backed company brain, this shape is what he meant. It's not a research project — it's the most practical, founder-friendly LLM-wiki variant.
Who this is for
- Founders who want one company brain in Markdown + git.
- Teams of 2–50 where institutional knowledge is scattered across people's heads.
- Companies with active AI adoption — the brain is updated by agents as they work.
- Anyone convinced that Markdown + git is the right long-term memory substrate.
Skip this if
You're a solo hacker with no team — gbrain's structure optimizes for organizational memory, not personal notes. Look at obsidian-mcp-tools or claudian instead.
What problem it solves
Every growing company has the same silent crisis: important knowledge lives in Slack threads, in the CEO's head, in old Notion docs nobody reads. New hires take weeks to ramp. Decisions get re-litigated because nobody remembers why they were made the first time. Context decays faster than you can write it down.
gbrain is the answer that doesn't require you to become a documentation-first culture. Your AI agents write the wiki as they work. You commit it to git. Every new hire clones the repo. Every AI tool reads from it. The brain is never someone's side project — it's the output of normal AI-assisted work.
How to install it (plain English)
- Clone or fork the gbrain repo. Pick a structure — most companies start with
/people,/projects,/decisions,/systemsat the top level. - Wire your agent to write to it. Your Claude Code / Cursor / other agent compiles session data into the right subfolders on a cadence (daily or weekly).
- Seed with what you already know. Write 10 People pages for your team. Write 5 Project pages for your current initiatives. The agent fills in the rest.
- Commit to git. Onboard every new hire to clone it.
Full walkthrough with prerequisites: /memory/tools/gbrain.
What you can do with it (for a non-technical founder)
- Onboard new hires in 1 day, not 3 weeks — they clone the brain, read it, ready to contribute.
- Never re-litigate a decision — grep the brain, find why you decided what you did.
- Audit what the AI thinks it knows — it's all Markdown. Read it. Fix wrongs. Commit.
- Survive vendor changes — if Claude becomes unaffordable or Cursor changes its pricing, the brain is portable. You own it.
- Build compounding context — every quarter's commits show what your company learned.
What CLO adds on top
gbrain gives you the content — the wiki of everything your company knows. Cognition CLO gives you the retention layer — which of your employees is actually retaining the content of which pages. The brain means your agents know your company; CLO tells you which of your humans are quietly forgetting the Q3 playbook before it costs you a deal.
Rule of thumb: gbrain is institutional memory. CLO is institutional retention.
FAQ
Who actually built gbrain?
The pattern originated with Karpathy's gist. Farza at F Inc built the first shipping implementation. Garry Tan (YC) popularized the company-brain flavor. gbrain as a distinct project took that flavor and productized it.
How is it different from using Notion?
Notion is a database; gbrain is a Markdown brain. Markdown is portable, git-versionable, AI-legible. Notion is proprietary and costs money to leave. For AI memory, Markdown wins.
Does my team need to use it directly?
Not necessarily. Your team writes normal work; your AI agents compile gbrain as they assist. If team members want to read or hand-edit, Markdown is the lingua franca.
How is it different from claude-memory-compiler?
claude-memory-compiler is Claude Code-native; it hooks into CC sessions specifically. gbrain is a structure (people/projects/decisions), agent-agnostic. You can point any LLM-wiki implementation at a gbrain-shaped directory.
Can I use it with Obsidian?
Yes. Point Obsidian at the gbrain directory and it renders like a standard vault with wikilinks. Best of both worlds.
What happens if the AI writes something wrong?
Edit the Markdown. Commit. The agent will read the corrected version next time. Auditable, editable, git-backed.
Ready to install? Full walkthrough at /memory/tools/gbrain. Comparison with obsidian-mcp-tools: /blog/gbrain-vs-obsidian-mcp-tools. Credit to @garrytan for popularizing the pattern and @FarzaTV for the original inspiration off Karpathy's sketch.
Share this post: