An LLM-maintained wiki is a persistent, interlinked Markdown knowledge base that a language model writes and keeps current. Each new source is read once and integrated — entity and concept pages updated, summaries revised, contradictions flagged — so the knowledge compounds instead of being rediscovered on every query. The canonical statement of the pattern is Karpathy’s LLM Wiki essay.
Against RAG
Retrieval-augmented generation retrieves raw chunks at query time and generates an answer; nothing accumulates, and subtle cross-document synthesis is re-derived every time. The LLM-maintained wiki inverts this: queries hit the wiki, where the cross-references already exist and the synthesis already reflects everything ingested. The maintenance burden that defeats human-kept wikis is near-zero for a model that doesn’t get bored.
Three layers, three operations
The architecture is immutable raw sources, the LLM-owned wiki, and a schema config that turns the model into a disciplined maintainer. The operations are ingest (integrate a new source), query (answer from the wiki, optionally filing the answer back), and lint (health-check for contradictions, stale claims, orphans, and gaps).
Lineage and kin
The pattern is the realization of Vannevar Bush’s 1945 Memex — associative trails between documents — with the maintenance problem finally solved. It is a sibling of Interpretable Context Methodology (the workspace shape of agentic work) and the agent-readable web standards (how the web exposes itself to agents): all are file-native, plain-text, and Markdown-first. It also extends the Semantic Web’s machine-readable ambition to a personal knowledge store. This site’s own wiki is an instance of the pattern.
🔍 Gap: The pattern is young — there is no published, longitudinal account of how an LLM-maintained wiki holds up over months of ingests without accumulating contradictions or drift. The lint operation is the proposed safeguard, but its effectiveness at scale is unproven.