Obsidian + LLM Wiki: A Second Brain, or a More Expensive Cache?
A workflow that lets the model maintain its own notes wiki; the tweet says it can save 80% of chat time; someone tested it on 1,914 pages of documents, and pure wiki scored only 0.64.
The core of this setup is connecting an Obsidian vault to Claude Code and then, following Karpathy's LLM Wiki pattern, letting the model maintain a wiki layer on its own. Mr. Buzzoni recently posted a tweet that combined Obsidian, Claude Opus 5.5, and Karpathy's LLM Wiki, saying that one page can turn a vault into a continuously running second brain. The accompanying video demonstrated this graph: 61 nodes, 95 links, and one hub.
First, some context. In llm-wiki.md, Karpathy describes a pattern: traditional RAG re-retrieves from the original documents on every question, so knowledge does not accumulate; LLM Wiki has the model incrementally maintain a persistent markdown wiki that sits between you and the raw materials. When new material comes in, the model reads it, extracts information, updates entity pages and topic summaries, and flags conflicts between new data and old conclusions. The wiki is a compounding asset, not something re-derived on every query.
Three-layer structure:
- raw sources: raw materials, immutable, read-only for the model.
- wiki: model-generated markdown pages—summaries, entities, concepts, comparisons, reviews—owned by the model.
- schema: tells the model how the wiki is organized and what the workflow is, such as CLAUDE.md or AGENTS.md.
There are three types of operations: ingest, query, and lint. During ingest, one source may touch 10 to 15 wiki pages; query answers can be archived back into the wiki; lint periodically checks for contradictions, outdated claims, orphan pages, and missing cross-references. There are also index.md and log.md—one manages the content catalog, the other manages the timeline. Karpathy's framing is that humans curate materials, ask questions, and judge meaning, while the model handles summarization, cross-referencing, archiving, and consistency maintenance.
Mr. Buzzoni's tweet gave several numbers: a new source reaches 11 pages within three hops; Claude reads those 11 pages and skips the rest of the vault; the top 10% of nodes hold 25% of the links, and a few hubs handle most of the routing. The tweet claims this setup cuts weekly chat time by 80% while burning only a small fraction of the quota. It also lists the tool stack: Obsidian, Karpathy's LLM Wiki, and claudesidian. claudesidian is a Claude Code + Obsidian starter kit with a preconfigured PARA directory, thinking mode and writing mode, several skills, Git version control support, optional Gemini integration for image analysis, and optional Firecrawl integration for fetching full web pages.
But the comments were not one-sided. A user named hallaai tested 1,914 pages of documents with 77 questions, including 12 FinanceBench 10-K/10-Q filings, investor decks, earnings reports, and 3 architectural drawings. Results: routing plus hybrid retrieval plus LLM reranking plus page-by-page visual transcription scored 0.89; stuffing the entire document into context scored 0.88; LLM Wiki plus its 5 most-cited source pages scored 0.79; classic hybrid chunk RAG scored 0.72; pure LLM Wiki scored 0.64. Of the 77 pure-wiki answers, 30 lost points, and 25 of those said the information was not in the wiki. The conclusion is that the wiki is more like a map and synthesis layer, not suitable as the sole source of truth.
Another user, hyunsoo36, said ingest costs are hard to recoup: the compiled wiki often lacks the details needed to answer, so the agent still has to go back and read the raw materials; the wiki becomes an extra layer of navigation and verification, adding LLM calls, tool calls, latency, and tokens. gimalay's usage is more restrained: numbers stay in CSV and JSONL source files, pages contain only links or queries, and the agent reads values rather than summaries; retrieval does not rely on index.md to select pages either, but instead queries frontmatter and link relationships, gets a short list, and then goes to the source files.
What this workflow really saves is bookkeeping: updating cross-references, keeping summaries fresh, flagging old-new conflicts, and maintaining consistency across dozens of pages. People usually abandon a wiki because the maintenance burden grows faster than the value. Models do not get tired and can edit 15 files at once, so maintenance cost is close to zero. The cost is losing detail during compilation, and the wiki potentially becoming an expensive cache. Before using it, think clearly: which pages in your vault are hubs, what will the model miss if it reads only 11 pages, and which facts must stay in the source files.
Tool links:
- Obsidian: https://obsidian.md
- Karpathy LLM Wiki: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
- claudesidian: https://github.com/heyitsnoah/claudesidian
- Original tweet: https://x.com/Mr_Buzzoni/status/2123364662255214776
发布时间: 2026-11-19 20:28