From messy pasted notes to an auditable, correctable, reusable second brain. This is the system diagram I drew after Karpathy's talk.
When your knowledge outgrows your memory,
you don't need more files, you need a navigable map.
Capture → transcribe → archive; wrong stays wrong. No loop, no correction, and more notes means more mess.
When AI uses the knowledge base it logs every miss, hallucination and stale fact, then feeds them back into the source files so notes correct themselves.
A 100-line Agent Skill that writes how well the AI worked back into the vault as revision records.
# skills/wiki-feedback/SKILL.md name: wiki-feedback trigger: "after every retrieval" on_hit: record(query, path, used=true) on_miss: record(query, reason="not-in-vault") on_wrong: record(query, correction, path) nightly: - aggregate misses → suggest new notes - aggregate wrongs → diff-patch old notes - commit to git, open PR for human review
Key optimisations · all shipped
Hallucination rate (vs. no feedback loop)
Single-query retrieval hit rate
189 were approved and merged by a human; 38 were rejected (data archived).
Clustered from miss logs, each with a traceable source.
No need to rewrite every note. Wire in one loop and every AI use improves your vault.
Blueprints, Skill and note templates are all at github.com/lewis/karpathy-wiki-v2