LLM wiki for your company: one wiki every AI reads

The pattern, in short

In April 2026 Andrej Karpathy published LLM Wiki, a short text on GitHub. It describes a pattern for building a personal knowledge base with an AI model. LLM stands for large language model, the kind of model inside Claude and ChatGPT. The pattern has three parts.

Your sources. Articles, papers, notes, transcripts. They stay as they are. The model reads them and never changes them.
The wiki. Plain text pages that link to each other: summaries, pages per subject, comparisons, an overview. The model writes them and keeps them current. You read them.
The rules. A short text that tells the model how the wiki is built, and how to keep it.

The model has three jobs.

Take in a new source. It reads the source, writes down what matters and updates every page the source touches. One source can change ten pages or more.
Answer a question. It answers from the wiki, and says which pages it used. A good answer can be kept as a new page, so what you found out stays.
Check the wiki. From time to time it looks for pages that contradict each other, and for claims that a newer source has overtaken.

The text is short on purpose. It describes the idea and leaves the details to you and your own AI tool. Some call what comes out an LLM knowledge base. It is the same thing.

Why it caught on

Most people know AI and documents as upload and ask. The model finds a few passages that seem to fit, and answers from those. That works. But the model works everything out again at every question, and nothing is kept.

An LLM wiki turns that around.

Knowledge is worked out once. Then it is read many times. The links between subjects are already there when you ask.
It gets better as you use it. Every source and every question adds to the pages.
You can check it. It is a wiki. Open a page and read what the model wrote.
The upkeep gets done. People give up on wikis because keeping one current is dull work. A model does not get bored, and it does not forget to update a link.

Your part is choosing the sources and asking good questions. The model does the filing. It suits anything where knowledge builds up over time: research into a subject, a book you are reading, your own notes.

LLM wiki or RAG

RAG stands for retrieval-augmented generation. Your documents are cut into passages and kept in a search index. At every question, the passages that seem to fit are looked up and handed to the model. The glossary defines both words in a few sentences.

An LLM wiki next to RAG
	LLM wiki	RAG
When the work is done	When a source comes in	At every question
What is kept	Written pages that link to each other	Passages in a search index
Can a person read and check it	Yes. It is a wiki	Only the source documents
How it grows	Each source and each question improves the pages	The index grows. The understanding does not
Where it fits	Up to a few hundred sources that matter	Very large or fast-changing collections
Used together	Holds what the company knows for certain	Reaches the long tail of documents behind it

A wiki is enough when the knowledge that matters fits in a few hundred sources. Karpathy's text says the pattern works at about a hundred sources and some hundreds of pages, without retrieval behind it.

Retrieval is the better tool for a very large collection, or for one that changes by the hour. Nobody writes a wiki page for every support ticket.

They combine well. The wiki holds what the company knows for certain. Retrieval reaches the long tail of documents behind it.

Where a personal wiki stops

Karpathy calls it a pattern for personal knowledge bases. He names a team as one place the idea could go, and leaves open how. For a team, a personal wiki stops where an AI second brain stops.

It has one owner. A colleague cannot read your wiki, and their AI tool cannot either.
It has no roles. Nobody decides who may change what, because only you change anything.
It has no shared history. When a decision changes, nobody else sees what changed, or why.
It lives in one place. One tool reads it, on one machine. The AI tools of the rest of the team do not.

What is a company brain? makes the same comparison for notes that an AI reads.

What changes when the wiki is a company's

People look for it as an LLM wiki for teams, for business or for companies, and as an enterprise LLM wiki. They mean one thing: the same idea, shared. Six things are added.

Roles decide who may change what. Administrators, editors and readers.
Every change has a history. Who changed what, and when. Every change can be undone.
Every AI tool reads the same wiki. Claude for one colleague, ChatGPT for the next, Copilot at work. One source for all of them.
Each answer carries its source. Anyone can check where it came from.
It stays current as you work. Meetings, mail and documents land in the same place. Correct something once, and it stays corrected for everyone who asks after you.
Agents act within limits. An agent that works from the wiki only gets the actions you switched on for it: agents that cannot do what you did not allow.

Dienox has its own name for a wiki like this: universal AI context.

How Dienox does it

Dienox is a platform that gives a company one knowledge base that every AI tool its people use reads from: Claude, ChatGPT, Copilot, Cursor and any other MCP client. If you know the LLM wiki, this is one for a whole company.

Start. A guided interview draws out what normally lives in people's heads. Half a day, and the knowledge base has its first version. Or start from the documents you already have. Dienox reads them, turns them into a first version, and lists the questions it could not answer.
Connect. Each person adds Dienox in their own AI tool as a connector, signs in and picks the organisation. About a minute per person, and nothing to install. The tools connect through MCP, the open standard for connecting AI tools to data and systems. The steps per tool are on connect your AI tools to your company knowledge.
Put it to work. On the same knowledge, agents take on recurring work: the weekly overview, inbox triage, meeting preparation. Every action has a limit.

Every AI tool you connect reads the knowledge base, and writes back what was decided. The home page shows the whole route: from a signup to a first agent.

When a wiki of your own is the better choice

Sometimes it is. You are one person. You like to build. You want everything on your own machine, and what you need to know fits in a few hundred sources. Then follow Karpathy's text. Hand it to your AI tool and build the wiki together. That is what it was written for.

It stops being the better choice when a second person needs to rely on it. Then you need roles, a shared history and a way in for every tool. Those three are the work. The glossary has the short version.

Questions people ask

What is an LLM wiki?

An LLM wiki is a set of linked pages that an AI model writes and keeps up to date from your sources. Instead of searching your documents again at every question, the model works the knowledge out once, writes it down and reads those pages afterwards. A person can open any page and check it.

Who came up with the LLM wiki?

Andrej Karpathy described the pattern in April 2026, in a short text he published on GitHub under the title LLM Wiki. He presents it as a pattern for building a personal knowledge base with an AI model, not as a product. The text is written to be handed to your own AI tool.

How do I build an LLM wiki?

For yourself: collect your sources, hand Karpathy's text to your own AI tool and let it write the first pages with you. Add sources one at a time, and read what it writes. For a team, building one also means roles, a shared history and a connection for every AI tool. That part is what Dienox provides.

Is an LLM wiki the same as RAG?

No. RAG, retrieval-augmented generation, looks up passages in a search index at the moment of each question and hands them to the model. An LLM wiki does the work earlier: when a source comes in, the model reads it and writes what it learned into linked pages. The first retrieves. The second writes things down.

Does an LLM wiki replace RAG?

Not always. A wiki is enough for the knowledge that matters most, up to a few hundred sources. For a very large or fast-changing collection, retrieval is the better tool. The two combine well: the wiki holds what the organisation knows for certain, and retrieval reaches the long tail of documents behind it.

Can a team share one LLM wiki?

Yes, with what a personal wiki lacks. The pattern was written down for one person, so a shared wiki needs roles that decide who may change what, a history of every change, and a way for the AI tool of every colleague to read it. Dienox provides those: one knowledge base for the whole organisation.

Do Claude and ChatGPT read the same wiki?

Not by themselves: each AI tool keeps its own projects and its own memory. With Dienox they do. Each person adds Dienox as a connector in the tool they use, signs in and picks the organisation. Claude, ChatGPT, Copilot, Cursor and any other MCP client then read the same knowledge base.

Is an LLM wiki a second brain?

It can be. A second brain is a personal system for notes, and an LLM wiki is one way to build it: the AI model does the writing and the filing, and you read. When a whole organisation shares the wiki, with roles and a history, it is a company brain rather than a second brain.

What happens to our wiki if we leave Dienox?

You take it with you. Your knowledge is kept in open files with its full history, and nothing is locked into a format only we can read. The same holds when you change AI tools: the knowledge base stays where it is, and the next tool you pick connects to it.