Andrej Karpathy's LLM wiki: notes your AI maintains for you

The problem with asking your documents

The usual way to give an AI your documents is retrieval: split them into chunks, find the chunks that look relevant to a question, and let the model answer from those. It works, but nothing is learned. Ask a related question tomorrow and the model reads the same chunks and works it all out again. Connections between two sources, or a contradiction between them, are found and then forgotten.

In April 2026 Andrej Karpathy, who co-founded OpenAI and led AI at Tesla, published a short note called LLM Wiki describing a different habit. Instead of searching raw documents at question time, the model reads each source once and folds what it learned into a set of linked Markdown pages. The understanding is written down, kept current, and builds on itself.

“The wiki is a persistent, compounding artifact.”
Andrej Karpathy, LLM Wiki

The note contains no code. Karpathy calls it an idea file: you hand it to an agent such as Claude Code or Codex, and the agent sets up the details with you.

Three layers

Raw sourcesraw/: Articles, papers, transcripts, data. You choose what goes in. The model reads them and never changes them, so there is always an original to check against.
The wikiwiki/: Pages the model writes: a summary for each source, a page for each person, product or idea, and comparisons worth keeping. You read it; the model owns it.
The schemaCLAUDE.md, AGENTS.md: The rules: where pages go, how they are named, what to do when a source arrives. You and the model refine it as you go.

Three things the model does

Ingest. You add a source. The model reads it, talks through the key points with you, writes a summary page, and then updates every page the source touches, often ten to fifteen of them. New facts go under the right heading with a citation; conflicts with what the wiki already says are marked rather than overwritten.
Query. You ask a question. The model answers from the wiki and cites its pages. If the answer took real work, such as a comparison or an analysis, it becomes a page too, so the next question starts from there.
Lint. Every so often the model checks the wiki's health: contradictions, outdated claims, pages nothing links to, ideas mentioned everywhere that still have no page of their own.

Karpathy prefers adding one source at a time and staying involved, so he can steer what gets emphasized. Batches work too.

Two files that keep it navigable

The index lists every page with a one-line summary, grouped by type. The model reads it first when answering, which, by Karpathy's account, is enough to find the right pages at around a hundred sources and hundreds of pages, without any vector search.

The log is an append-only record of every ingest, question and lint pass, one entry per line with a date and a fixed prefix, so it can be searched and so the model can see what happened last time.

Why it works now

The idea is old. In 1945 Vannevar Bush imagined the Memex, a personal store of documents joined by trails of associations. Personal wikis have been possible for decades. They rarely last, and Karpathy names the reason:

“Humans abandon wikis because the maintenance burden grows faster than the value.”
Andrej Karpathy, LLM Wiki

Updating fifteen pages after reading one article, fixing broken links, noticing that two sources disagree: that bookkeeping is exactly what a language model does well and people do not. So the work splits cleanly. You pick the sources, ask the questions and decide what matters. The model summarizes, cross-references and files.

“Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.”
Andrej Karpathy, LLM Wiki

Running it on notespace

Karpathy's setup is a folder of Markdown files on his computer, browsed in Obsidian and edited by a coding agent. notespace keeps the same shape but stores the wiki on a server that every AI app you use can reach: Claude, ChatGPT, Grok or a script. The built-in karpathy-llm-wiki methodology lays a space out like this:

raw/YYYY-MM-DD-<slug>  the source, verbatim, locked when saved
wiki/sources/<slug>     one summary per source
wiki/entities/<name>    people, companies, products, places
wiki/concepts/<name>    ideas, methods, terms
wiki/analyses/<slug>    answers worth keeping
wiki/index              every page, one line each
log                     what was done, newest last

Sources stay original. The server locks every note under raw/ when it is saved. No assistant can rewrite a source, even by mistake; only you can unlock one, from your account page.
Pages say what they are. The server rejects a wiki/ page without a type (source, entity, concept or analysis), so the wiki stays sortable as it grows.
One ingest, one change. An ingest that touches fifteen pages is a single batched call that applies completely or not at all, so a dropped connection never leaves the wiki half updated.
Links are tracked. [[links]] between pages are indexed, so the lint pass can list broken links and pages nothing points to, and moving a page rewrites the links to it.
Search past the index. When the wiki outgrows its index, keyword and semantic search (Chinese included) take over. Karpathy suggests adding a separate search tool at that point; here it is built in.
Nothing is lost. Every edit keeps the previous version, and you can export the whole space as a zip of Markdown files, ready to open in Obsidian.

Start one

Connect notespace to your AI app, then ask in plain words. Your assistant proposes the change and nothing happens until you confirm it.

You

Set up my research space as a Karpathy LLM wiki.

Then, whenever you read something worth keeping

Ingest this article into the wiki. What are the key points?

And later

What does my wiki say about evaluation methods? Where do the sources disagree?

Changing a space's method never edits notes that are already there. If the rules do not fit how you work, ask your assistant to draft your own and approve them the same way.

Source: Andrej Karpathy, LLM Wiki (GitHub Gist, 4 April 2026). Quotations are from that note.