An LLM wiki is a knowledge base that an AI model writes and maintains itself. You drop raw sources into a folder, the model compiles them into linked pages with summaries and backlinks, and it periodically checks its own work for contradictions and gaps. Andrej Karpathy popularized the pattern in April 2026 as a simpler alternative to retrieval-augmented generation for his own research.
It works well for one person and one project. A company version is a much harder problem, and it's the one I spend most of my time on.
How Karpathy's LLM wiki works
As VentureBeat described it, the setup has three parts:
- Raw sources. Papers, repositories, datasets and web articles go into a raw folder, converted to markdown.
- Compilation. The model reads that material and writes a structured wiki: summaries, concept pages and backlinks between related ideas.
- Health checks. The model regularly scans the wiki for inconsistencies, missing information and new connections, and fixes them.
Everything stays in plain markdown, so a person can read, edit or delete any page and trace every claim back to a file. Karpathy reported that at around 100 articles and 400,000 words, the model navigates the wiki well using summaries and index pages, with no retrieval infrastructure at all.
LLM wiki vs. RAG
| LLM wiki | Traditional RAG | |
|---|---|---|
| How knowledge is stored | Readable markdown pages written by the model | Document chunks turned into vector embeddings |
| How the model finds things | Index pages, summaries and backlinks | Similarity search over the embeddings |
| Can a person audit it? | Yes, page by page | Hard; vectors aren't readable |
| How it improves | The model rewrites and links pages over time | Only when documents are re-indexed |
| Sweet spot | One person or project, hundreds of documents | Large document sets |
Retrieval-augmented generation still makes sense for large collections, usually backed by a vector database. The LLM wiki's real contribution is the idea that the model should organize knowledge ahead of time instead of searching raw files every time it's asked something.
Ontology, world model, company brain, LLM wiki: same idea, different scale
Meanwhile, Palantir's CEO has been telling everyone that the ontology is what makes AI useful in an enterprise, and other large platforms are building their own versions. The labels multiplied, but they describe one category: a company's data, organized with its context, that people and AI agents can both use.
The differences are mostly scale and emphasis:
- Ontology: a structured model of the entities, relationships, events and rules in a business. Customers connect to deals, deals to projects, projects to tasks, tasks to people.
- World model: the system's understanding of what's happening in the business right now, what has changed and what matters.
- Company brain: the friendlier small-business version. SOPs, history and decisions, made searchable.
- LLM wiki: a file-based layer an LLM can navigate easily. The smallest and simplest version.
Why the personal version breaks at company scale
Commentators on Karpathy's post suggested a company-wide version, and he agreed it could be a new product category. Here's what changes when you try it.
The sources never stop. A researcher adds papers when they choose. A company generates thousands of emails, messages, call transcripts and CRM updates a day. Nobody is going to clip those into a folder.
Knowledge goes stale. Processes shift, people change roles, goals move. I've seen plenty of AI systems work well for six months and then quietly stop being used because the context they were built on no longer matched how the business ran.
Much of what matters is never written down. Every company runs on unspoken rules: which client needs careful handling, how a certain team actually makes decisions, what a thumbs-up in Slack means. That knowledge lives in conversations, not documents.
Not everyone can see everything. A personal wiki has one reader. A company has permissions, and so should every agent that reads from it.
The system has to know when not to act. A useful company knowledge layer isn't just a retrieval engine. When it isn't confident, or an action is risky under company policy, it should hold back and route to a person rather than guess its way into production.
What a company LLM wiki actually needs
The version we build starts with events rather than documents:
- Capture events at the source. Emails, meetings and messages are captured as they happen, not exported later.
- Add context. For each event, work out what it is, who's involved, and which account, project or case it belongs to.
- Store it in a structured layer. That can be a knowledge graph, a relational database, a vector store or a hybrid. The structure matters more than the technology.
- Make it retrievable by people and agents, with permissions applied.
- Apply confidence rules. Low-confidence matches and risky actions go to a human in the loop.
Because it updates from live events, it stays current without anyone maintaining it, and it improves as it learns the business's entities, relationships and recurring workflows. A sales agent drafting an email can see the full account history first. A support agent can see past tickets on the same issue, including which answer actually prevented a refund.
Why this matters more than which model you pick
The AI layer on top is becoming interchangeable. Models from different labs keep closing the gap on each other, and the chat products built on them are converging too. What can't be swapped out is your business: the history, decisions and tribal knowledge underneath.
That's why I'd tell any business owner to focus on getting their company context into a layer they own before worrying about which AI tool to standardize on. Buying everyone a chat subscription and teaching them to use it is useful, and we do plenty of it. But on its own it doesn't build anything you keep. If your context lives in your own layer, you can point any model at it and switch whenever a better one ships. This is also what most people mean by context engineering, done once at the company level instead of rebuilt for every agent.
How to get started
If you want to understand the pattern, build a personal LLM wiki first. A folder of markdown, a model with file access and an hour of setup will show you how it works.
For a company, don't start by exporting everything. Pick one type of record your team works every day, like accounts, cases or projects, and one event source, usually email. Capture, add context and store that one flow well before adding the next.
FAQ
Who came up with the LLM wiki?
Andrej Karpathy described the pattern in April 2026, calling it an LLM knowledge base. He uses it to keep research knowledge persistent across AI sessions.
Is an LLM wiki better than RAG?
For a single person or project with hundreds of documents, it's often simpler and easier to audit. For large or constantly changing collections, retrieval systems or a structured knowledge layer still make more sense.
Do you need a vector database for an LLM wiki?
Not at small scale. Karpathy's version uses markdown files, summaries and index pages instead of embeddings.
Is an LLM wiki the same as a company knowledge base?
It's the smallest version of the same idea. A company knowledge base for AI also needs automatic capture from email, chat and calls, permissions, and rules for when the system shouldn't act.