# A wiki that maintains itself: the Karpathy LLM-wiki pattern

Source: https://lodos.md/en/blog/self-maintaining-wiki-karpathy-pattern
Published: 2026-06-22 · Author: Samet Samyeli · Language: en · Reading time: 3 min

Most company wikis rot the moment you stop tending them. The Karpathy LLM-wiki pattern turns a knowledge base into something your AI keeps alive. Here's how.

---

Every company wiki starts with good intentions and ends as a graveyard. Pages drift out of date, links break, half-written drafts pile up, and eventually nobody trusts it enough to look. The knowledge is technically *there*; it's just quietly rotting.

The **LLM-wiki pattern**, popularized by Andrej Karpathy, offers a way out: stop treating the wiki as a static archive and start treating it as something a language model continuously reads, answers from, and repairs. lodos builds its living wiki directly on this idea.

## Answers, not searches

A traditional wiki gives you a search box and wishes you luck. An LLM-wiki gives you an answer.

Ask "how do we handle Stripe webhooks?" and the agent doesn't dump ten pages on you. It narrows to the relevant ones with full-text search, synthesizes a reply from your **own verified pages**, and cites each source as a `[[wikilink]]` so you can trace every claim back to where it came from.

> The difference between a guess and an answer is a citation. The wiki only ever answers from pages you've written and verified, never from the model's imagination.

## It keeps itself fresh

The reason wikis rot is that maintenance is nobody's job. The LLM-wiki pattern makes it the system's job.

A daily scan, cheap SQL and regex, no AI cost, flags the rot before it spreads:

- **Orphans**: pages nothing links to.
- **Dead links**: `[[wikilinks]]` pointing at pages that no longer exist.
- **Stale claims**: pages that haven't been touched while the things around them changed.
- **Gaps**: questions that get asked but have no page to answer them.
- **Unfiled pages**: notes that never made it into the structure.

You wake up to a short, honest health report instead of discovering the decay months later.

## It grows as you work

The best part is that the wiki gets richer without anyone sitting down to write it. Every decision the AI records, every refund it processes, every task it finishes gets written back into your knowledge base. The wiki compounds as a side effect of running the company.

## A graph the AI can walk

Underneath the pages is a knowledge graph. Every `[[link]]` is resolved by one shared engine, the same one used for rendering, for linting, and for the agent's own reasoning.

That last part matters. When the AI answers a multi-hop question, "which customers are affected by the contract change we decided on last quarter?", it walks the graph the way you would, following the thread from customer to contract to decision. Crucially, it traverses **structure only**: slugs, titles, and the edges between pages. It never needs the page bodies or any secret to understand how your business fits together.

And because the graph is Obsidian-grade and built in, you can import a real Obsidian vault, keep editing in lodos, and export back to plain markdown any time. No lock-in.

## Links that never break

There's one small architectural decision that makes the whole thing durable: **slugs are stable, titles are free to change.** Rename a page however you like, as often as you like; every link to it and every edge in the graph stays intact. A wiki that breaks when you rename a page is a wiki people stop maintaining.

## Why it lives locally

All of this runs on your machine. The pages, the graph, the daily lint, the agent that reads them: none of it ships your knowledge to a server. It's the same principle behind the rest of lodos: the AI gets the context it needs, and your data stays where it belongs. We wrote more about that in [why we built a zero-knowledge company OS](/blog/the-zero-knowledge-company-os).

A wiki shouldn't be a chore you eventually abandon. It should be a map of your company that gets sharper the more you work. That's what the LLM-wiki pattern makes possible, and what lodos ships in the box.

[Explore the product](/product) or [download lodos](/download) to start your own.
