Karpathy 的 LLM Wiki pattern

Terminology rule (zh-TW pages)

技術名詞首次出現以「中文 (English original)」格式呈現，例：嵌入 (embedding)。不自創翻譯——若無公認譯名直接保留英文（如 RAG、 wiki、schema）。代碼、API 名、CLI flag、套件名、檔名一律不翻。

整理 Andrej Karpathy 在 2025 年底提出的「由 LLM 維護的個人知識庫 (personal knowledge base maintained by an LLM)」pattern。這頁是文件, 不是 skill——先把想法存下來,之後再決定要 vendor 既有實作（如 obsidian-second-brain）還是自己寫一個 local skill。

來源

• gist.github.com/karpathy/442a6bf555914893e9891c11519de94f ——長版「LLM Wiki」pattern 文件,設計成可以直接貼給你自己的 LLM agent, 讓它跟你共同設計具體實作。
• x.com/karpathy/status/2039805659525644595 ——原始 tweet（"LLM Knowledge Bases"）,gist 的~一頁濃縮版,也是整個 pattern 的種子。

TL;DR

Wiki 是一個持久 (persistent)、會累積 (compounding) 的產物。 RAG 每次 query 都從 raw sources 重新 derive 知識;LLM Wiki 把該知識 compile 一次進結構化 markdown,並隨新來源 ingest 持續維護更新。

Obsidian 是 IDE,LLM 是 programmer,wiki 是 codebase。 ——Karpathy, LLM Wiki gist

原始 tweet（逐字英文引用）

LLM Knowledge Bases

Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So:

Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them.

IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides).

Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale.

Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base.

Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into.

Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries.

Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows.

TL;DR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

架構 (Architecture)

三層,出自 gist:

1. Raw sources ——你篩選過的源文件（articles、papers、images、 datasets）。Immutable——LLM 只讀不改,這是 source of truth。
2. The wiki ——LLM 生成的 markdown 目錄（summaries、entity 頁、 concept 頁、comparisons、overview、synthesis）。LLM 完全擁有這層, 你讀,LLM 寫。
3. The schema —— CLAUDE.md / AGENTS.md,告訴 LLM wiki 的結構、 慣例 (conventions)、ingest / query / lint 的 workflow。你跟 LLM 一起 隨時間共同演化 (co-evolve) 這份 schema。

三個 operations

Operation	做什麼
Ingest	把新 source 丟進 raw/,LLM 讀完、寫 summary 頁、更新對應 entity / concept 頁、更新 index、append log。一個 source 可能 touch 10–15 頁。
Query	提問。LLM 先讀 index,drill 進相關頁,然後回答——可以是 markdown 頁、比較表、Marp 簡報、matplotlib 圖。好答案要回填 (file back) 進 wiki,explorations 才會累積。
Lint	定期健檢 (health check)。找矛盾、stale 主張、orphan 頁、缺 cross-reference、資料缺口。並建議下一步要 investigate 的問題跟要找的新 sources。

兩個特殊檔案

• index.md ——content-oriented 目錄。每頁一條 link + 一行 summary, 按類別分組。LLM 每次 ingest 都更新,query 時先讀。在 ~100 sources / 數百 頁規模下,不需要 embedding-based RAG 基礎建設就能 work。
• log.md ——chronological,append-only。建議用統一 prefix（如 ## [2026-04-02] ingest | Article Title）讓 log 變 greppable: grep "^## \[" log.md | tail -5。

為什麼 work

殺掉個人 wiki 的不是讀或想,是維護 (maintenance)。更新 cross-reference、 保持 summary 最新、標出新舊矛盾、跨數十頁維持一致性——人類會放棄,因為 維護負擔成長得比價值快。LLM 不會無聊、不會忘了補 back-link,而且一次能 touch 15 個檔案。Wiki 持續被維護,因為維護成本接近零。

這個 pattern 精神上接近 Vannevar Bush 的 Memex （1945）——一個私人、主動策展 (curated) 的知識庫,文件之間的關聯 (associative trails) 跟文件本身一樣重要。Bush 沒解掉的部分是誰來維護。 LLM 解掉了。

周邊生態 (Ecosystem)

圍繞這個 pattern 長出來的實作跟相關工具:

• eugeniughelbur/obsidian-second-brain ——一個 Claude Code skill（不是 Obsidian plugin）,把 pattern 推得更遠: 31 個 slash commands + 排程 agents + Python scripts。加進 AI-first 慣例（machine-readable preamble、bi-temporal facts——同時 track「事實 何時為真」+「vault 何時學到」、Two-Output Rule——每個答案都順便更新 相關頁面）、vault-first research（先掃既有 notes 再上網查）、 _CLAUDE.md / index.md / log.md / SOUL.md / CRITICAL_FACTS.md （~120 tokens 永遠載入）放在 vault root。 "If Karpathy's wiki is a knowledge base you maintain with an LLM, this is a knowledge base that maintains itself."
• tobi/qmd ——本地 markdown 搜尋引擎 （hybrid BM25 + vector + LLM re-ranking,全部 on-device）。同時有 CLI （給 LLM shell 出去用）跟 MCP server。Gist 提到當 index 檔擋不住時的 自然升級路徑。
• Obsidian Web Clipper、Marp（markdown-to-slides）、Dataview （對 frontmatter 跑 query）——gist 推薦的支援工具。

跟本 repo 的關係

本 repo 的 project-knowledge-harness 是任務記憶 (task memory) —— TODO.md + backlog/ + pitfalls/, 圍繞已交付/延後工作跟過去踩到的坑。Karpathy 的 LLM Wiki 是知識記憶 (knowledge memory) —— 對某個外部研究領域策展、合成 (synthesize) 後的 view。兩者正交 (orthogonal),可以在同一個 project 共存。

可能的後續（未承諾）:

• Vendor obsidian-second-brain 進 skills/vendor/（license 允許的話）, 跟 project-knowledge-harness 並列當「兩種 memory pattern」的對照組。
• 寫一個極簡 local skill llm-wiki-bootstrap ——只做 gist 描述的三層 架構 + index.md / log.md + ingest / query / lint 三個指令, 不綁 Obsidian。

如果要正式追蹤,用 scripts/add-todo.sh 加一條 P3。