Obsidian CEO Open-Sourced His Personal Notes Vault. But a Second Brain Isn’t for ‘Storing’
Obsidian CEO Steph Ango's personal vault template has been open-sourced for free on GitHub, with 7 repositories ready to use. The real highlight is a paradigm discussion around the second brain: Karpathy's LLM Wiki, rvaniaaa's compilation architecture, and the community's disagreement over cost and hallucination issues.
Steph Ango, the CEO of Obsidian, uses the GitHub username kepano. He has open-sourced his personal vault template at github.com/kepano/kepano-obsidian. A vault is the workspace concept in Obsidian; put simply, it is a local folder where all your notes live.
This vault is not a demo project; it is the setup he uses every day. Download it, unzip it, and open it in Obsidian, and you have a second brain you can start using right away, locally. The README is direct too: ‘No dogma—take the parts you need.’ People have already starred the repositories on GitHub, but starring does not equal using. Most people will probably still stop at the blank page.
The seven repositories released this time cover the entire chain from getting started to daily maintenance:
- kepano-obsidian: The CEO's personal vault template, with a complete directory structure, solving the problem of opening to a blank page
- jsoncanvas: An open canvas format; .canvas files are plain text, so your canvas data always belongs to you
- obsidian-sample-plugin: The official plugin template, written in TypeScript; if you want to build plugins, start here
- obsidian-minimal: A minimalist theme made by kepano, freeing the vault from the default skin
- obsidian-clipper: The official web clipper; one click in the browser and the page goes straight into the vault
- obsidian-importer: Migrate old notes from other apps, so you don't have to start from scratch
- obsidian-help: A plain-text version of the official documentation; reading the docs is reading Markdown
Repositories are tools. What is more worth looking at than the tools is the shift in thinking taking place in the Obsidian ecosystem.
rvaniaaa posted an article on X with a very long title: The Second Brain Is Not a Storage System. It's a Compiler. The article cites the LLM Wiki concept that Andrej Karpathy proposed in April 2026. At the time, the concept spread quickly on GitHub: 5,000 stars, and a post about folder structure got 16 million views. One sentence Karpathy wrote in llm-wiki.md has been quoted repeatedly:
RAG re-derives knowledge on every query. A compiled wiki derives it once and keeps it current.
RAG re-derives knowledge on every query. A compiled wiki derives it once and then keeps it up to date.
rvaniaaa translated this into a concrete architecture. Three folders, one file, one loop. raw/ is only an input buffer; the brain itself is not here. wiki/ is where compilation happens: when a source comes in, the model reads it, extracts from it, links it to existing pages, and flags contradictions. CLAUDE.md is a dynamically maintained personal profile that the model reads before every session. Once the whole loop is running, when you open your computer, the compilation has already been done in the background, and you continue from where you last stopped thinking. Retrieval is answering questions; compilation is accumulating understanding.
This direction makes sense, but the discussion in the comments is more clear-eyed than the article.
Mike Leiter raised a cost issue: RAG does not cost money at query time; the token cost of a compiled wiki happens during the ingest stage. For high-frequency, high-value knowledge, compilation is absolutely worth it. But for a random document, if it is compiled into a wiki page and never retrieved again, that is sunk cost. The balance point for a second brain lies in trade-offs: invest more in high-frequency, important content, and don't build a wiki page for every paper.
echoes2099 worried about another problem: once files reach the hundreds, ‘putting new files in the right place’ starts to become difficult. Models hallucinate unless the entire process is designed very explicitly. Northmeadow added two safeguards: every derived page must point to its source, and when the source changes, the page must be marked stale; contradictions must be flagged for humans to handle and must not be quietly merged. Otherwise, one bad source will rewrite fifteen pages before you notice.
These discussions themselves are evidence that the direction is taking shape. Tools are already being built in the comments, such as akm (Agent Knowledge Manager), a capability-library manager for coding agents that can directly index Karpathy-style LLM wiki formats (schema.md + raw/ + pages/). Tools are chasing the ideas.
Templates, tools, and methodology are all in place. The final step is for you to get it running yourself. Feed the things in raw/ into it and watch it compile.
发布时间: 2026-11-22 02:37