Karpathy’s LLM Wiki: Build Your Own Knowledge Base

Andrej Karpathy recently shared a simple but powerful pattern for managing knowledge with LLMs. It is not a product or a library — it is an idea file: a single Markdown document you paste into any coding agent (Claude Code, OpenAI Codex, pi/OpenCode, …) and let the agent build and maintain a personal wiki for you.

• Original gist: LLM Wiki
• Walkthrough this post is based on: Andrej Karpathy’s LLM Wiki

The core idea: stop retrieving, start compiling

Most “chat with your documents” tools are RAG: they search your files at query time and reassemble an answer from fragments every single time. Nothing accumulates.

Karpathy flips this: the LLM compiles your raw sources once into a wiki and keeps it updated.

The wiki is a persistent, compounding artifact: summaries, cross-references and contradictions are already there. You rarely write the wiki — the LLM does the bookkeeping.

“Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.”

Three layers

1. Raw sources — your curated PDFs, articles, notes. Immutable: the LLM reads them but never edits them. They remain the source of truth.
2. The wiki — a folder of Markdown files (entity pages, concept pages, summaries) owned entirely by the LLM.
3. The schema — an AGENTS.md (or CLAUDE.md) file that tells the agent how the wiki is structured and how to ingest, query and lint it.

Three operations

• Ingest — drop a source, and the LLM reads it and updates 10–15 pages: a summary, the index, entity pages, links, and any contradictions.
• Query — ask a question; the LLM reads the already-synthesized wiki and answers with citations. Good answers get filed back as new pages, so they are not lost in chat history.
• Lint — a periodic health check: contradictions, orphan pages, stale claims, missing cross-links.

Two files that hold it together

• index.md — a catalog of every page (link + one-line summary). The LLM reads it first to find relevant pages before drilling in.
• log.md — an append-only history of ingests and queries. Use a consistent prefix such as ## [2026-09-24] ingest | Title so grep "^## \[" log.md | tail -5 shows the latest activity.

Build one in minutes

mkdir llm-wiki && cd llm-wiki
mkdir raw

1. Open your agent in that folder.
2. Paste the gist and say: “Set up an LLM Wiki here. First ask me what it is about and what sources I will feed it, then write the schema.”
3. Answer a few questions (topic, source types, page types).
4. Drop a source in raw/ and say: “Ingest raw/bitter-lesson.pdf” — the agent generates the summary, updates the index, and creates linked pages.
5. Ask synthesis questions such as: “How do these two authors agree and disagree?”
6. Open the folder in Obsidian and look at the graph view.

Tools

• Obsidian — the viewer and graph: see your wiki as a network of linked notes instead of a flat folder.
• Obsidian Web Clipper — a browser extension that saves any web article to Markdown, straight into your raw/ folder. The fastest way to feed new sources.

RAG vs LLM Wiki

• RAG — best for huge, changing corpora and precise chunk citations. Stateless: every query starts from scratch.
• LLM Wiki — best for a curated corpus (roughly 100–500 sources) and deep work where synthesis matters more than retrieval. Stateful: knowledge compounds.

One real risk: a wrong summary can get “baked in” as fact and quietly propagate across linked pages. Mitigate it with the lint step and an occasional spot-check against the raw source.

Why it matters

The hard part of a knowledge base was never reading or thinking — it is the bookkeeping: links, summaries, keeping claims current. LLMs do that for free, which is why humans usually abandon wikis. It also makes Vannevar Bush’s 1945 Memex vision finally practical: the connections between documents become as valuable as the documents themselves.