It only spoke when spoken to

A session of mine ran a backup command that quietly excluded a folder from protection. My notes held the exact opposite decision: that folder was a deliberate archive, the whole point was that it must never be excluded. When I later searched the memory with the command as the query, the contradicting note came back at similarity 1.0. A perfect match. It had been sitting there the entire time, and nothing looked.

Nothing looked because I had been choosing when to look. The memory spoke at session start, in a briefing. It spoke when specific tools fired specific curated gates: a web search, a file that was already ingested, a handful of command patterns I had thought to list. The long middle of a working session, the hundreds of shell commands and edits where the actual decisions happen, was silent territory. My curated list was a bet on where nuance would appear, and the incident landed squarely outside the bet, as incidents do.

The reframe, once I said it out loud, was embarrassing in its simplicity. Selecting when to consult the memory is a prevention posture: it catches the categories of mistake I already predicted. But only the memory itself can know whether it holds something relevant to this moment. If it only participates at moments I preselected, I have not built an amplifier. I have built a checklist with extra steps.

So now every tool call is embedded and matched against the whole index, every command, every edit, all day. That part turned out to be cheap; the warm index answers in a fraction of a second. The expensive part is deciding when to speak, because at any reasonable threshold the memory always has something vaguely relevant to say, and a colleague who comments on everything is a colleague you stop hearing.

I ran it in shadow mode for days, logging what it would have said without saying anything. At the naive cutoff it would have spoken on half of all tool calls. Desensitization as a service. The fix came from the calibration work the library had just audited: score each match against that note’s own history, weight in how novel the note is and how often its past appearances were actually useful, and let one composite number decide.

Then I went to pick the threshold and stopped myself, because I was about to hardcode a magic number into a system whose whole point is that the data should decide. The nightly calibration now derives the bar from the trailing week’s score distribution, floored so a quiet week cannot lower it into noise. The morning I shipped it, the derivation came back with 65.1. The number I had been about to pick by hand was 65.

It speaks a few times a day now, unprompted, when something it holds genuinely bears on what is passing through my hands. Whether the judgment is good enough to trust is next month’s question; the telemetry that will answer it is already flowing. What changed this month is smaller and bigger than that: the memory no longer waits for permission to be relevant.