llm-wiki

A personal, compounding tech knowledge base maintained by an LLM. Based on Andrej Karpathy's LLM-Wiki pattern.

What Is This?

Traditional AI assistants re-derive knowledge from scratch on every query — expensive, slow, and non-cumulative. LLM-Wiki is different.

You feed it sources (articles, docs, papers, notes). An LLM compiles them into a persistent, interlinked knowledge base in a structured wiki format. Future queries search the wiki — not the raw internet, not raw documents — resulting in fast, token-efficient answers that compound over time.

The more you ingest, the smarter and cheaper the system becomes.

How It Works

Three zones, strict access rules:

/query never reads from raw/. It always reads from the compiled wiki/.

Directory Structure

Why Each Directory Matters

**vocabs/** — The most important directory after wiki/. Every wiki page uses canonical terms from here. This prevents vocabulary drift (same concept with 5 different names) and enables the tag-based search in wiki/index.md to work reliably.

**wiki/sources/** — Every ingested article gets a source page. This is the bibliography. When a wiki concept page says [[martin-fowler-cqrs]], you can trace the claim back to its origin. /lint uses this for contradiction detection.