The Weekly Triage Pass
A register that only grows is a register nobody reads. Twenty minutes a week keeps it honest.
Any register an agent writes to will grow until it stops being read. A weekly pass with four verdicts — merge, keep, kill, promote — plus a hard cap that forces the pass when breached, keeps it usable. The rule that makes it survive: nothing is deleted without a reason and a pointer.
Prerequisites
At least one register the agent appends to: patterns, findings, ideas, open questions. A cap on how many live entries it may hold.
If nothing appends automatically, this is premature — hand-written registers rarely reach the size where triage pays.
Steps
1. Set a hard cap first. Ours is 30 live entries. The cap is not about storage; it is a forcing function. A register with no cap grows until reading it costs more than the value of any entry in it, and that point arrives without announcement.
2. Run the pass on a fixed day. Weekly, same slot, attached to something that already happens. A triage that has to be initiated will not be.
3. Give every entry over 7 days one of 4 verdicts:
merge — duplicate of an existing entry, fold it in
keep — still live, still relevant
kill — no longer relevant, archive with a reason
promote — act on it now; produces a task, not a note
4. Cap promotions. At most 3 per pass. An unlimited promote turns triage into a planning session, which is a different activity that will crowd it out entirely.
5. Archive with a reason and a pointer. Nothing is deleted outright:
killed 2026-07-27 — superseded by the build-level check; see [entry]The reason is what stops the same observation being re-added three months later by someone who cannot see why it was dropped.
6. Add a tripwire. A daily count of entries older than a threshold, with a flag if it exceeds a limit. Not new work — an alarm, so the cap is not the only thing standing between the register and neglect.
7. Keep the archive searchable, not just present. Killed entries go somewhere they can be found by the same search that covers live ones. The value of an archive is answering did we already consider this — and an archive nobody can search answers no better than deletion.
8. Log the pass itself. One line: date, entries reviewed, verdict counts. Two months of those lines show whether the register is stabilising or the cap is being defended by increasingly aggressive killing, which is a different and worse equilibrium than a register that has stopped growing.
Verify
The register is healthy when the pass takes under 20 minutes and the count is stable rather than climbing. A pass that keeps taking longer means either the cap is too high or the entry gate is too loose.
Check the kill rate. Zero kills over several passes means the pass is not doing its job — a register where everything stays relevant is a register nobody is honest about.
Troubleshooting
The pass keeps getting skipped. It has no executor. Attach it to an existing recurring task rather than relying on the day of the week.
Everything gets kept. Killing feels like losing information. The archive-with-reason step exists precisely to make it feel safe; if it is being skipped, the reluctance is rational.
The register was falsely empty for weeks. Check the tooling before congratulating anyone — a parser reading the wrong column produces a clean empty queue and no error at all. That specific failure ran 21 days here before anyone opened the file by hand.