LLM-maintained wiki¶

Definition¶

The LLM-maintained wiki is the pattern, popularised by an April 2026 gist of Andrej Karpathy's, in which an LLM incrementally builds and maintains a persistent wiki as its knowledge substrate — writing and revising pages as material arrives — instead of retrieving over raw documents each session.

Explanation¶

The pattern targets a compounding problem: the more one reads, the less one retains, and every LLM session starts from zero. Retrieval-augmented setups answer this by searching source documents at query time, but the sources stay unprocessed — nothing accumulates. The wiki pattern moves the work to write time: the model distils incoming material into named, linked pages and revises them as understanding changes, so the knowledge base compounds and any future session (or human) reads consolidated pages rather than re-deriving from raw sources. It is a pattern, not a tool — the staged source is third-party coverage of the gist, and the primary artifact remains the better reference for specifics. The trade against retrieval is explicit: maintenance cost and the risk of distillation errors are accepted in exchange for persistence, legibility, and curation — the wiki is inspectable and correctable in a way an embedding index is not.

Key Properties¶

Distillation happens at write time; knowledge accumulates in named, linked, revisable pages
Sessions read consolidated pages instead of re-retrieving raw sources
Human-legible and correctable, unlike an embedding index
Accepts maintenance cost and distillation risk in exchange for compounding

Relationships¶

Retrieval-Augmented Generation (RAG) — positions itself as the write-time alternative to RAG's query-time retrieval: distil into maintained pages once rather than search raw sources every session
Ebbinghaus forgetting curve — supplies the mechanism behind the problem the wiki pattern solves — unreviewed knowledge decays on this curve, so distillation into revisable pages replaces re-reading
Context layer — contrasting substrates for agent knowledge: the wiki distils at write time into curated pages, the context layer indexes sources at read time and layers curation on top

Applications¶

Personal or team knowledge bases maintained by an agent across sessions; giving a long-running agent durable memory whose contents a human can audit and correct.

Sources¶

https://www.aibuilderclub.com/blog/karpathy-llm-wiki