Company brains / 23 Jun 2026

How to build a company knowledge base your AI maintains

Andrej Karpathy wrote recently about how he has started building personal knowledge bases with a language model. He drops source documents into a raw folder, and then he has the model compile them into a wiki of plain markdown files. Summaries, concepts, articles, links between them. The part that stuck with me is that he rarely touches the wiki himself. The model writes it and the model maintains it. He is the source and the editor. The machine is the author.

I have spent a year arguing that every business needs a knowledge base a machine can read. This is the next move. You stop writing the knowledge base by hand, and you let the model compile it for you.

The job nobody finishes

Most companies have tried to write things down. They have a wiki, or a shared drive, or a help centre. It started well and then it rotted.

The reason is always the same. Writing documentation is a job, a real one, and it is nobody's actual job. So it gets done in a burst when someone is new or bored, and then the company keeps moving and the document stops. Six months later it describes a business that no longer exists.

The knowledge did not go away. It scattered. It moved into inboxes, into Slack threads, into spreadsheets, into the head of the one person who remembers why you stopped using that supplier.

The work keeps producing knowledge.
Nobody has the time to keep writing it down.
So the written record falls behind the real one, and stays behind.

The inversion

Here is the shift Karpathy is pointing at. The thing that made documentation rot was that a person had to write it. Take that person out of the writing and the rot stops.

You still gather the raw material, because only you have it. The past quotes, the policies, the supplier emails, the support replies, the decisions and the reasons behind them. You drop all of it, messy and undated, into one folder.

Then the model compiles. It reads the pile and writes the wiki. It summarises each source, pulls the repeated ideas into concept articles, and links them to each other so the thing is navigable. When you add new raw material, it recompiles. The wiki is not something you keep up to date. It is something the model regenerates from the truth you keep feeding it.

What "compile" actually means

It helps to be concrete, because "let the AI write your wiki" sounds like magic and it is not. It is a loop with four plain steps.

Ingest. A raw folder where every source lands in whatever form it arrived. A pasted email, a clipped web page, a PDF, a screenshot, a voice note someone transcribed. No structure required. Getting it into the folder is the only discipline.
Summarise. The model reads each source and writes a short, clean note that says what it is and what matters in it. This is the index, and it is what makes the rest cheap.
Concept articles. The model notices that nine of those notes are really about returns, or about one carrier, or about a single product line, and it writes a proper article for each concept, drawing on all nine.
Link. It connects the articles to each other and back to the sources, so a person or an agent can walk from any fact to the things around it.

None of this needs a fancy retrieval system to start. At the scale most companies live at, the model maintaining its own index and summaries is enough to find what it needs. You reach for heavier machinery later, if ever.

The data has to be clean enough to compile

The model is good at this, but it is not a magician. Feed it three rate sheets that contradict each other and it will faithfully compile the contradiction.

So the unglamorous step still matters. Before the wiki is worth querying, someone has to drag the scattered knowledge into the open and clean it. The good news is that the same model that compiles the wiki is the one that helps you clean the raw pile. It flags the duplicates, the stale rates, the policy that two documents describe two different ways. You are not cleaning alone any more.

This is the same point I keep coming back to. Making the work legible is most of the job. The compiling is the easy, almost free part that sits on top.

Why a company should care

A self-compiling knowledge base is not a tidier wiki. It is a different asset.

It stays current, because it is regenerated from raw material instead of edited by hand.
It survives people leaving, because the knowing is in the pile, not in their heads.
It is the thing every agent you build reads from. The support agent, the quoting agent, the onboarding agent. They are all only as good as the brain behind them.

The point is not to have documentation. The point is to have one living source of truth that both your people and your machines can read, and to stop paying a person to keep it alive by hand.

Prompt you can use

I want to turn a pile of messy company material into a knowledge base you maintain for me, instead of me writing it by hand.

Here is how I want to work:
1. I will describe the sources I have: rate sheets, support replies, policies, supplier emails, past decisions, whatever it is. Ask me to list them.
2. For each source, write a short, clean markdown summary note that says what it is and what matters in it.
3. Once you have the summaries, tell me which concepts show up again and again across them, and draft a proper article for each concept that pulls from every relevant source.
4. Link the articles to each other and back to the sources, and propose a simple folder structure.
5. Flag anything that contradicts itself, looks stale, or is missing, so I can fix the raw material rather than bake the error in.

Start by asking me what sources I have and what this business actually does.

The rule

Stop trying to write the company wiki. You will never finish, because the work that produces the knowledge never stops.

Keep one raw folder of the truth. Let the model compile it into something legible. Maintain the pile, not the prose. The business with the richest brain wins, and the only sustainable way to keep that brain current is to stop maintaining it by hand.