A second brain that runs itself

Every tool I’ve used to get organized eventually became a place where data went stale: the app I stopped opening, the dashboard nobody checked, the second brain with no pulse. The problem was never storage. The problem was that nothing lived there.

So I built one that runs itself. My notes are a git repo of plain markdown, and every layer of it is agentic. A dozen scheduled AI routines clone the repo, do their work, and push back. By the time I sit down in the morning, the inbox is sorted and the daily note is written. I didn’t ask. That’s the point.

The operating system frame

The thing clicked when I stopped thinking of it as a notes app and started thinking of it as an operating system. Markdown files are the file system. Scheduled agents are the processes. One file, CLAUDE.md, is the kernel: the conventions, the voice, the policy every agent reads before acting.

The frame isn’t decoration; it tells you what to build. Operating systems have memory hierarchies, so the vault got one: raw capture flows into an inbox, agents distill it into evergreen notes, and the durable lessons get promoted into the kernel itself. Operating systems have schedulers, so attention became a scheduled resource: the system decides each morning what deserves to be surfaced, and warns me the day before an overloaded day arrives, not the morning of.

Agents propose, I accept

The rule that makes the whole thing trustworthy: agents propose, I decide. The gardener agent drafts new notes but can’t create them; its proposals show up on a dashboard where accepting one is a single tap. The curator agent watches my daily retrospectives for lessons that keep recurring and proposes edits to the kernel the same way. The system literally suggests amendments to its own constitution, and I ratify them with my thumb.

This matters because an agent that writes directly into your notes is a roommate; an agent that proposes is a chief of staff.

The system catches itself

My favorite part is the eval harness. It scores every agent’s output against a deterministic rubric: did the daily note fill every section, did it name specific files, did it follow the style rules. No judge model, just checks that mean the same thing every run.

Its first run caught something perfect: the daily agent had been breaking the vault’s own no-em-dash style rule every single morning, in every note, for weeks. The kernel said one thing; the process did another; the eval made the gap visible. A system that watches itself honestly beats a system that promises to behave.

What I’d tell you to steal

The files are the system. Whatever you build, keep the state in plain text you own, in a format that outlives any tool that touches it. The agents are stateless on purpose: everything important is a file in git, which means the whole brain is portable, diffable, and recoverable.

And give it a pulse. A second brain you have to push is a filing cabinet. One that runs ahead of you, sorts its own inbox, and shows up sharper than yesterday is something closer to a colleague.