LLM Wiki

An LLM wiki (Andrej Karpathy’s term) is a knowledge base the model itself maintains: structured, interlinked markdown pages. When a new source arrives, the model integrates it rather than storing it whole: it reads it, extracts what matters, updates the entity pages it touches, revises the topic summaries, and notes where the new material contradicts what the wiki already claims. The wiki is a persistent artifact that compounds; the sources stay immutable underneath it.

Three layers keep it honest: the raw sources, untouched; the wiki the model writes; and a schema that defines what pages exist and how they are maintained.

I run my reference corpora this way: a glossary index at the top, wikilinked pages below, and ingest that means integrate, not append. Reading is then navigation: the model opens the index, follows the links it needs, and loads nothing else. Agent memory at its best is the same pattern, notes the model maintains and revises across sessions rather than a transcript it searches. The public-web version of the idea is llms.txt, a curated markdown index a site offers so a model reads structure instead of scraping HTML.