Karpathy's LLM Wiki Compiles Knowledge

This is less a product than a strong operating model for long-horizon agent memory: compile once, maintain continuously, and keep the artifact inspectable by humans. The interesting part is not “better search,” it’s disciplined knowledge bookkeeping.

• –It shifts the bottleneck from retrieval to maintenance, which is where real knowledge systems usually fail.
• –The workflow is strongest in domains with steady source intake: research notes, team docs, due diligence, and personal knowledge bases.
• –The contradiction and lint loop matters more than embeddings here; provenance and stale-claim detection are what make the wiki trustworthy.
• –The human stays in the loop as source curator and reviewer, while the LLM handles the tedious cross-linking and updates.
• –If this becomes a product category, the moat will be around ingestion, review, and governance workflows, not just another vector search layer.