Wink Pings

Obsidian CEO Open-Sources His Second Brain: It's Not a Storage System, It's a Compiler

kepano put his personal vault template on GitHub, with 7 repositories available for free download. rvaniaaa's article breaks down a more fundamental issue: a second brain should work like a compiler, not be used as a filing cabinet. Most people got it wrong from the very beginning.

Obsidian's CEO (known as kepano on GitHub) has put his personal vault template on GitHub. The entire second brain, free to download. You can just copy it, open Obsidian, and stop staring at a blank page.

That's a generous move. But what's more worth pondering is this: most people, even with this template, will still end up wasting it.

In the article "The Second Brain Is Not a Storage System. It's a Compiler.", rvaniaaa cites the LLM Wiki concept Karpathy proposed in April 2026 and points out a key distinction. Karpathy's original words are:

RAG re-derives knowledge on every query; a compiled wiki derives it once and keeps it continuously up to date.

Most people treat their second brain as a filing cabinet. Store notes, search when needed, and hope the right things surface on their own. That's retrieval. Retrieval has a ceiling.

Compilation mode is different. When a source comes in, the system reads it, extracts key points, connects it to existing content, updates related pages, flags contradictions, and archives it. The next time a source is added, it builds on the already-compiled understanding. The library grows larger, and the compiler grows smarter.

Architecturally, it's just three folders plus one file: raw/ is the input buffer, wiki/ is where compilation happens, and output/ is built from the compilation results. At the center is CLAUDE.md, a dynamic profile maintained by the model and read at every session.

Why compile instead of store? Maintenance burden. A personal wiki quietly rots. It's not that the owner doesn't care, but that maintenance costs grow faster than value accumulates. Karpathy put it bluntly:

Collecting information is effortless. Keeping fifty interlinked notes up to date, consistent, and cross-referenced is work no one can sustain long-term.

Compilation transfers the burden to the model.

The 7 repositories released by the Obsidian team and CEO this time happen to cover the infrastructure of this system:

  1. kepano-obsidian: The CEO's own vault template, a direct copy of the entire second brain
  2. jsoncanvas: An open canvas format, where whiteboard data is saved as plain files and always belongs to you
  3. obsidian-sample-plugin: The official plugin template, start here if you want to write a plugin
  4. obsidian-minimal: A clean theme that makes your vault look sharp
  5. obsidian-clipper: Official web clipper, save any page directly into your vault
  6. obsidian-importer: Migrate old notes from other note-taking apps, no need to start from scratch
  7. obsidian-help: Official documentation, in plain text form, ready to consult anytime

The article also acknowledges the limitations of this approach. Quality depends entirely on source quality. Bad sources entering the compiler will contaminate every compiled page, far more dangerous than bad sources entering a filing cabinet. The first few weeks will also feel slow, with a small graph and obvious connections; value only emerges after 50 to 100 sources have been compiled. And the loop depends on the paid version of Claude Desktop, so it doesn't work on the free tier.

Karpathy's llm-wiki.md got 5,000 stars within a few days, and a post about folder structure got 16 million views. The signal is clear: the focus of knowledge management is shifting from "what's stored" to "what the system has digested for me."

The CEO's vault is already open source. The compiler can start running. The rest is to stop treating yourself as a filing cabinet.

发布时间: 2026-11-22 06:06