Karpathy's LLM Wiki challenges standard RAG

The LLM Wiki represents a shift from "search-as-memory" to "documentation-as-memory" where persistent markdown storage avoids vector database lock-in and provides a human-readable audit trail. Its "linting" mechanism allows agents to proactively identify contradictions and stale claims, creating a self-healing knowledge base. Early implementations like llm-wiki-compiler and CacheZero report up to 70x reduction in token usage by querying condensed summaries rather than raw chunks, mirroring human research workflows that build value over time.