llm-wiki

A persistent, LLM-maintained knowledge base, built on the llm-wiki pattern popularized by Andrej Karpathy. Instead of re-retrieving and re-reasoning over raw documents on every query (classic RAG), an LLM agent incrementally compiles your sources into a structured, interlinked wiki of Markdown files. Knowledge compounds over time instead of evaporating between sessions.

This repository is a scaffold. Folders and template files contain [PLACEHOLDER: ...] blocks marking exactly where your own content goes. Replace them as you ingest real sources.

The three-layer model

• Raw is your source of truth: PDFs, articles, notes, transcripts, datasets, PRs.

You drop files in; you never hand-edit them.

• Wiki is everything the agent writes: summaries, concept pages, entity pages,

comparisons, an index, and a change log. Humans mostly read this layer.

• Schema (CLAUDE.md) tells the agent how the wiki is structured and which

workflows to run. It is the single most important file — it turns a chatty model into a disciplined wiki maintainer.

Repository layout

Each wiki/ subfolder ships with a _template.md showing the expected page shape. Delete the templates once you have real pages, or keep them as a reference.

Quickstart

• Pick your agent. This scaffold assumes Claude Code,

which auto-loads CLAUDE.md. For OpenAI Codex or other agents, see AGENTS.md.

• Read & adapt the schema. Open CLAUDE.md and fill in the