My notes rewrite themselves every night

A self-hosted knowledge system that distils each day into a linked wiki my assistants can cite. What it does, what it costs, and where it still annoys me.

In March I spent two days rebuilding a decision my team had already made in November. The reasoning existed. I had written it down at the time, in a note I never opened again, in an app I had stopped using by February.

That was the fourth or fifth time. Capture was never my problem. I have thousands of notes. The problem is that a note only pays off at the moment, months later, when it changes what you decide, and by then you have forgotten it exists.

So I stopped treating the pile as an archive and started treating it as an input.

The setup

Everything runs on a Mac mini in my apartment. That is a privacy decision, not a budget one: the vaults hold my work, my plans and conversations I would not paste into someone else’s service.

Inputs land as plain markdown in two git repositories, one personal and one professional, kept strictly apart. Calendar events, tickets, meeting notes, saved conversations. Nothing clever, just files.

At half past two each morning a job wakes up, reads whatever arrived that day, and updates a wiki. One page per person, per project, per decision, per topic. The important part is that it rewrites those pages rather than appending to them. A page about a project is not a log of everything ever said about it. It is what is true about that project this morning, with the contradictions resolved and the stale parts removed.

Links between pages go into a Postgres table, so I can ask which decisions came out of which project, or what a particular person has touched. When I ask a question, the answer comes back with links to the pages it drew from. If the answer is wrong, I fix the page, and every future answer improves.

That last loop is what makes the thing worth running. Retrieval over a pile of notes gives you the notes back. Rewriting gives you a summary that gets better while you sleep.

The part I did not expect

The professional vault grew a section I never designed. Each feature I have built or led ended up with its own page: what it was, what my role was, what shipped. I did not sit down to write a portfolio. The nightly pass did it as a side effect of reading my tickets.

When review season came around I had the material already, with dates and links. That saved me a weekend and made the writeup more honest than the version I would have reconstructed from memory.

What it costs

Two things, and both are real.

The synthesis runs on a mid-tier model every night and a stronger one once a week to find contradictions across pages. That weekly pass is where most of the spend goes, and it is the pass I would keep if I had to drop one.

The other cost is trust. A system that rewrites your own words can rewrite them wrong. Markdown in git is what makes that survivable: every pass is a commit I can diff and revert. I read those diffs for the first two months. I check them monthly now, and I have reverted three.

If you build one

Start with the loop that reads, not the tools that capture. Capture is solved and every app does it. The read side is where these systems die.

Keep the vaults separated. Mixing personal and work knowledge feels efficient for about a week, until an assistant answers a question about a colleague with something from your therapy notes.

Use plain files. No database lock-in, no export path to maintain, and the whole thing outlives whichever tool you are using this year.

I now build these for other people, mostly founders who have the same complaint I had. The pitch writes itself, which is a warning sign, so here is the honest version: it took me four attempts over two years to get one that I still use.