Second Brain
A living record of how I think — maintained by an AI, owned in plain text.
Stats for Nerds:
- Team Size: 1 (+ Claude as the maintainer)
- Timeline: Ongoing
- Language: Markdown
- Framework(s):
- Obsidian (vault + graph)
- Claude / Cowork (the maintaining agent)
- Karpathy LLM-wiki pattern + Matuschak evergreen notes
- Tools:
- Wiki-links, INDEX hub notes, append-only logs
- Auto-memory, scheduled tasks, live artifacts
- Stave: synthesis-brain on Glyph
- Repository: Personal
The Concept
What is it?
It’s a knowledge base I interface with almost entirely through conversation — I don’t manage the files by hand. I talk to Claude naturally: sharing ideas, asking questions, doing research. Claude captures what’s worth keeping into the right note automatically, and reads from the vault at the start of relevant conversations to bring past thinking forward. Over time it becomes a living record of how I think, and Claude is the primary interface to it.
The whole thing is plain Markdown. That’s the point — the vault is the asset, the AI is interchangeable. If a model has an outage, prices spike, or a better agent shows up, the brain ports by swapping one instructions file. Notes that live inside a proprietary database aren’t a portable knowledge asset; these are.
The Architecture
How it's organized
Two patterns do the structural work.
Karpathy’s raw-vs-synthesized split. Source material lives in a raw layer the AI reads but never rewrites; the AI maintains a synthesized layer of concept and project notes on top. Three operations carry the system — ingest (process a new source, update the wiki, append to the log), query (ask the synthesized layer, not the raw pile, because synthesis already happened), and lint (a periodic health check for orphans, broken links, and stale claims).
Matuschak’s evergreen notes are the thinking-quality layer. Each concept note holds a single atomic idea, is named for the idea itself rather than its source, and links densely to everything it touches. When the same idea shows up in two separate notes, that’s the trigger to factor it out into its own concept — and the backlinks then reveal cross-project connections for free.
INDEX files act as Maps of Content — you read the hub note first and follow it down, so finding things stays fast even as the vault grows.
The Synthesis
The engine
A lifetime audit of the vault showed the real workflow isn’t ingesting clipped articles — it’s converse → decide → build → log. So the raw layer here is conversation itself, and “the Synthesis” is how that becomes knowledge, running at three speeds:
- Live, every conversation — a decision gets written to its project log the moment it’s made; a recurring idea gets extracted to a concept now; every add updates an INDEX. The vault log line is the receipt, never a to-do.
- Daily — a light evening pass verifies the day’s work landed where it belongs.
- Weekly — a Sunday triage tells the week’s story, runs a safety-net sweep for anything live capture missed, then rotates the log into an immutable archive.
The health metric is simple: capture misses per week should trend toward zero. A clean sweep means the brain grew correctly in real time.
Status
Ongoing and in daily use — it runs my projects, my reading, even my chess notes. The vault is where the thinking actually lives; the Garden is just where some of it surfaces.