karpathy-llm-wiki-implemented

An implementation of Andrej Karpathy's LLM Wiki pattern — compile your sources into a persistent, interlinked wiki that compounds over time instead of retrieving from raw files on every question.

The core idea

Most AI setups work like this: you upload files, ask a question, the model retrieves chunks and answers. Every question starts from zero. Nothing builds up.

This vault works differently. When you drop a source in — a PDF, a web clip, a transcript — you ask the agent to ingest it. The agent reads it once, extracts what matters, and merges it into a structured wiki: a source summary page, updated concept pages, entity pages, cross-links, and a note of any tension with claims already in the wiki. The next time you ask about that topic, the wiki already has the answer compiled and linked. You are not paying to re-read the same paper again.

The wiki gets richer every time you add a source or ask a question. Cross-references are already there. Contradictions are already flagged. Synthesis reflects everything you have read. The LLM's job is the maintenance work — the summarising, cross-referencing, filing, and bookkeeping that humans abandon. Your job is to curate sources and ask the right questions.

Architecture: three layers

**raw/ is immutable.** The agent reads from it but never writes to it. These are your source of truth.