本文由 AI 分析生成
建立時間： 2026-04-09 來源： https://x.com/defileo/status/2042241063612502162
Summary
A viral guide presenting Andrej Karpathy’s “LLM Wiki” pattern — using Claude Code as the writer/maintainer of a persistent, compounding Obsidian knowledge base. The core idea differs from RAG: instead of retrieving from raw documents at query time, the LLM incrementally builds and maintains a structured interlinked wiki, integrating every new source into existing pages with cross-references, contradiction flags, and evolving synthesis. The wiki is a persistent artifact that compounds; the human curates sources and asks questions.
本文介紹 Karpathy 的「LLM Wiki」模式：LLM 不是 RAG 式查詢原始文件，而是增量式地建立並維護結構化的 Obsidian wiki，每次新增來源時更新實體頁、修訂主題摘要、標記矛盾。人類負責策展來源與提問，LLM 負責所有書目工作（摘要、交叉引用、歸檔）。
Key Points
- Three-layer architecture: raw sources (immutable), wiki (LLM-generated markdown), schema (CLAUDE.md defining wiki conventions and workflows)
- Ingest operation: one source → LLM extracts, writes summary page, updates index.md, updates 10-15 cross-linked concept/entity pages
- Query operation: LLM reads index, drills into relevant pages, synthesizes answer with citations — best answers get filed back into wiki as new pages
- Lint operation: weekly health check for contradictions, orphan pages, stale claims, missing cross-references
- index.md is content-oriented catalog (the LLM reads this first for any query); log.md is chronological append-only record
- Key difference from RAG: knowledge is compiled once and kept current, not re-derived on every query — cross-references and synthesis already exist
Insights
The Vannevar Bush Memex analogy (1945) is apt: Bush envisioned private, curated knowledge stores with associative trails. What he could not solve was who does the maintenance — the LLM handles that. The observation that “creativity requires intuitive recombination of knowledge in ways not obvious enough to be amenable to intentional retrieval” explains why parametric in-weights knowledge differs from harness-based memory in kind, not just degree. This connects directly to the Defining Continual Learning arguments about automaticity.
Connections
Raw Excerpt
“The wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you’ve read.”