An overnight AI pass over a day's notes, capped and checked
Dreaming
Part of Life Tracker
A nightly job that asks a small AI model whether pairs of the day's notes are related, duplicated or in conflict, and leaves a short report by morning. The model has no tools and a spending cap, and any answer that does not quote the notes exactly is thrown away.
Problem
Related notes drift apart, duplicates pile up, and two notes can disagree about the same fact. I wanted an overnight pass that finds these without giving a model free rein over my records while nobody is watching.
Approach
Each night Dreaming pairs notes that look related and asks a small model whether each pair is related, a duplicate or a contradiction. It changes nothing unless I switch writing on, and even then only two labelling fields and three kinds of one-line fix, each undoable. Everything else is a flag or a proposal in a morning report. It never deletes, renames, moves or sends anything.
Architecture
A scheduled job starts it early each morning. The model it asks is a small one, called with no tools at all and a spending cap on every call. It is deliberately not a full agent run, which might still be going when the day starts. Each night has a fixed budget, usage is checked before every call, and anything over the budget waits for the next night. It stops in time for the report to land before the morning email.
Whatever the model says is treated as data, never as an instruction. Notes reach it marked as material to compare, so a note worded like an order is compared rather than obeyed. Every verdict must quote a complete line from each note exactly, or it is thrown away, and the code checks every quote against the file. Only the scheduled job is allowed to call the model or change a note. A night is two passes: the one that may write reads only text I or the system wrote, and a second reads outside material such as received mail and can only propose.
What I built
With writing switched on, it makes only a few changes a night, each checked against the file as it was read and logged before and after, and it leaves alone any note someone else has edited recently. Undo restores a field only if it still holds what Dreaming wrote, then pauses that rule for a week. The report is kept short, and its headline reaches the daily email.
Results
Replayed against two weeks of tasks I had actually closed, the proposed rule for closing them automatically would have closed none, and matching my own words in session logs to open items was rarely right. So the first version reported rather than edited. The first scheduled night ran on time and produced a handful of proposals and no changes. Writing has since been switched on, and the night of 13 October made ten changes alongside 34 proposals and no conflicts.
What didn't work
Most of what a night reads produces nothing yet. On the test night, most of the files it gathered led to no output, and what to do with them is a decision for the next version.
Stack
- Python
- Claude Code
- Claude Haiku
- launchd
- git
How these figures are sourced
The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from the morning report of 13 October. Every figure and claim on this page is tied to a source file or a CV line in the site's content.