# Karpathy's LLM Wiki Pattern — Original Reference Source: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f ## Core Insight "The wiki is a persistent, compounding artifact. The knowledge is compiled once and then kept current, not re-derived on every query." Human curates sources and asks questions; LLM maintains the knowledge system. Obsidian becomes the IDE, the LLM becomes the programmer, and the wiki becomes the codebase. ## Why This Beats RAG Traditional RAG rediscovers knowledge on every query — it searches raw sources, pulls relevant chunks, and synthesizes an answer from scratch. The LLM Wiki compiles knowledge once into maintained pages, so queries hit pre-synthesized, cross-referenced content. ## Key Operations | Operation | What it does | When to use | |---|---|---| | **Ingest** | Read new sources, extract key information, update 10-15 wiki pages, maintain consistency | When new documents arrive | | **Query** | Answer questions against compiled wiki with citations | When the user asks something | | **Lint** | Identify contradictions, orphaned pages, stale claims, missing cross-references | Periodic maintenance | ## Recommended Tools - **Obsidian** — IDE for browsing and exploring the wiki - **Web Clipper** — Browser extension for converting articles to markdown - **Marp** — Markdown-based slide decks from wiki content - **Dataview** — Obsidian plugin for querying page metadata - **qmd** — Local search engine with BM25/vector hybrid search ## Applications - Personal tracking (goals, psychology, self-improvement) - Research (building comprehensive understanding over weeks/months) - Book annotation (companion wikis with characters, themes, plot connections) - Team/business (wikis from Slack threads, meeting transcripts) - Due diligence, competitive analysis, trip planning ## Community Extensions Worth Knowing - **Provenance tracking** — Record which source files produced each claim, detect staleness through content hashing - **Hierarchical inheritance** — Parent-child page relationships instead of flat indexing - **Decision records** — Capture why the wiki evolved, not just what changed - **Two-tier LLMs** — Local models for sensitive data, cloud for the rest - **Graph databases** — Typed ontologies instead of markdown links