Building a Self-Maintaining AI Wiki with Karpathy’s Pattern

ow to implement a persistent knowledge base that compounds over time using LLMs for automated bookkeeping.

The Evolution of RAG

Traditional RAG systems rediscover knowledge from scratch on every query. The LLM Wiki pattern introduces a persistent, compounding artifact.

Human-in-the-loop: Humans curate sources while the LLM handles the “grunt work” of filing and summarizing.

Incremental Ingest: New sources are integrated into existing pages rather than just indexed.

Compounding Value: Cross-references and contradictions are flagged during ingestion.