Blog
How to build a company brain, part 1: what it is and why AI agents need one
- Company brain
- AI agents
- Knowledge management
An AI agent starts every task knowing nothing about your company. So you paste the same background in again, and it still guesses the parts you forgot to mention. A company brain is my fix for that: one folder of plain files that every person and every agent reads before doing anything.
I run three of them: one for my company, one for my own life admin, and one for a product team I build with. This series explains how they work, in ten parts. I can't show you the real ones, so every example uses a made-up company: Kranich Studio, 14 people in Berlin, who build software for clients and sell one product, an appointment-booking app called Termino.
Key takeaways
- A company brain is a folder of linked Markdown files plus a short list of written rules. People and agents read it before they act.
- Agents file, summarise and update it on a schedule. People approve, send, pay and decide.
- Raw sources are kept as they arrived. Current knowledge and the decision log sit on top, and every page links back to where it came from.
- You need Markdown, one SQLite file and, later, a knowledge graph. The files stay on your machine, but the agent reading them usually runs on a cloud model.
What a company brain is
It's the company's memory, written down so that an agent can read the relevant part in seconds. There's an entry page that says where everything lives. There are dated logs of what happened, folders for clients, money, deadlines and open work, and a short list of rules every agent has to follow.
Say the founder of Kranich Studio asks an agent where invoice INV-2026-041 stands. The agent doesn't need a briefing. It reads the entry page, then the finance folder's summary, then the note for that invoice, and answers with a link to the evidence. A second agent planning the Termino roadmap starts from the same pages, so the two aren't working from different versions of the company. A new hire can read them too.
If you know llms.txt, the idea is close. That proposal gives a website one short Markdown file with some background and links to the detailed pages, written for a model. The entry page does the same job for a company's own files.
What it is not
It isn't a wiki. A wiki depends on someone remembering to update it, and it goes stale the week that person gets busy. A brain is kept current by routines. Every evening an agent goes through new mail, calendar entries and documents, files what arrived and writes the day's log. A person reads the summary and decides what to do about it.
It isn't a chatbot either. The brain is the material, and agents are the clerks who read and file it. You can change the agent or the model and keep the brain.
The layers
All three of my brains keep raw material apart from what is true now. Sources are stored exactly as they arrived, with a summary on top. Journals hold the daily and weekly digests. Work folders hold one project or feature each, with versions that freeze once they're approved. Knowledge pages describe how things are today. The decision log records what was decided and why, and nobody rewrites it.
The technology is plain. Markdown holds the content, because people can read it and so can any language model. SQLite, a database that lives in a single file on your disk with no server to run, makes it searchable through its full-text index. A knowledge graph, which I build with Graphify, shows how the notes connect.
One caveat about the word "local". The files, the index and the graph sit on your own machine. The agent that reads them usually sends what it reads to a cloud model, and Graphify also uses a model when it reads documents (code it parses without one). That's why part 7 is about what an agent is allowed to see.
What this series covers
You can read any part on its own. In order, they go like this:
- Part 2: the benefits, including knowledge that stays when people leave.
- Part 3: use cases in operations, personal admin, product research and sales.
- Part 4: how to build it, with the folders, the summary files and a one-week plan.
- Part 5: collaboration, which means sharing it through GitHub with branches and pull requests.
- Part 6: skills, the step-by-step jobs every agent does the same way.
- Part 7: the rules, and how to keep personal data out.
- Part 8: research, done and filed so it can still be trusted later.
- Part 9: scaling, or why big brains get slow and costly and how a knowledge graph helps.
- Part 10: the SQLite index, a searchable index that your agent builds and maintains.
Frequently asked questions
What is a company brain?
A folder of linked Markdown files plus written rules that people and AI agents read before acting. Agents keep it current on a schedule, and people approve anything that matters.
Do I need special software?
No. You need Markdown files in a private Git repository and an AI agent that can read and write files, such as Claude Code or Cursor. A small SQLite index comes later. The files and the index live on a laptop. The agent's model usually runs in the cloud.
How is it different from a wiki?
A wiki waits for someone to update it. A brain is updated by routines that agents run every day, and it carries rules that tell agents how to read it and how to write to it.
Sources
- The llms.txt proposal llmstxt.org
- About SQLite (single file, no server) sqlite.org
- SQLite FTS5 extension documentation sqlite.org
- Graphify on GitHub github.com
Talk to me
Building an AI product and need someone senior to own the technical side? Book a free 30-minute call.