Summary: A knowledge-base pattern published by andrej-karpathy (in April, reaching ~20M views) in which an LLM, not the human, maintains the knowledge: it stores, connects, updates, and checks for contradictions, so the value compounds the more you use it. Three layers: raw sources kept unchanged as the source of truth, a wiki of digested concepts and links, and a schema (often just a CLAUDE.md/AGENTS.md) with the rules for ingesting, using, and maintaining it. Karpathy suggested it fits up to ~100 sources; beyond that, retrieval needs more care.
Key ideas
- Raw is “the dump”: transcripts and clippings, rarely accessed directly; the wiki is the clean, digested layer (osim-tochna—second-brain-and-llm-wiki, amit-bendor).
- Its biggest value is the connections between pieces of information (osim-tochna—second-brain-and-llm-wiki, dana-maman).
- Your own outputs count as sources too: notes, summaries, conversation logs, articles you wrote. Otherwise the AI won’t cite them (osim-tochna—second-brain-and-llm-wiki, dana-maman).
- Start simple per Karpathy (sources / wiki / schema) and restructure freely as needs emerge; Dana restarted several times (osim-tochna—second-brain-and-llm-wiki).
- Rated suitable for ~100 sources; past that, see wiki-retrieval (osim-tochna—second-brain-and-llm-wiki).
- Team adoption: monday’s Harmony team applied the pattern to a 20-person org. The key difference from RAG is persistence: a librarian who remembers yesterday’s question. Useful answers (e.g. a competitor pricing table a PM researched) are written back, so the next question needs no re-research (startup-for-startup—353-self-updating-team-brain).
- Their schema: concepts, domains, entities (companies, people by role), decisions, and action items, all interlinked; one meeting can update 15 pages. Raw items are stored so nothing is re-processed (saves tokens) (startup-for-startup—353-self-updating-team-brain).
- Advice: start with basic, manual, local versions to prove value, before cloud CI and scale (“bicycle before car”) (startup-for-startup—353-self-updating-team-brain).
- Singapore’s foreign minister published his own NanoClaw fork with a Karpathy-style LLM wiki for memory, using it to draft speeches, answer questions, and research as a “second brain” (hidden-layers—nanoclaw-gavriel-cohen, gavriel-cohen).
- Team context for agent teammates: boards, docs, GitHub, and the team’s public and private channels sync to a file system, an LLM wiki runs on it, and agents consume it, alongside memory at the agent and task-session level (ai-engineering-podcast—atlas-ai-teammate, netanel-abergel)
Disagreements & open questions
- File layout: Dana keeps Karpathy’s flat three layers; Amit experimented with a tree by domain (research, engineering…) aimed at fast, correct lookup. Both agree it keeps getting reorganized (osim-tochna—second-brain-and-llm-wiki).
Takeaways
- After any restructure, ask the AI to verify system integrity: links and routines still work (osim-tochna—second-brain-and-llm-wiki)
Related
second-brain · wiki-retrieval · open-knowledge-format · knowledge-rot