Why LLM-Maintained Wikis Could Replace RAG

Andrej Karpathy published a gist proposing a three-layer knowledge architecture - raw sources, an LLM-generated markdown wiki, and a schema file like CLAUDE.md - as a replacement for traditional RAG pipelines. The core idea: instead of re-deriving context per query, let the LLM maintain a persistent wiki where cross-references and synthesis compound over time through ingest, query, and lint operations. Each new source updates 10-15 wiki pages. The tooling is deliberately local - Obsidian as the wiki IDE, markdown files for full control, BM25/vector hybrid search via qmd. Knowledge bases have always failed because maintenance falls on humans. Karpathy's bet is that LLMs finally absorb that bookkeeping cost, making the wiki pattern viable again. The approach maps cleanly to how teams are already using CLAUDE.md and AGENTS.md files as living project context.