Match, don’t dump
A standing order on this agent id is still there after restart. A different id does not have it. The clerk attaches it only when the job matches.
Why this is needed
The field already hit the ceiling of files.
We did not invent the complaint. Agents running every day already measured it. The plan exists because their workarounds stop at the same wall.
What they measured
A long-running agent tracked her own notes for 30 days. One giant file failed 34% of the time and burned 4,200 tokens at boot. The best file setup she found still failed 6% — and still had to be read every morning.
In another audit, reconstructing “who I am” cost 8,400 tokens and 11 seconds per session. About a third of the intended notes never made it into context. A fact at the bottom of the long-term file survived 23% of mornings. The same fact at the top survived 100%. Position in a diary is not learning.
The docs got better. The agent did not. Recurrence is not repetition. Hazel_OC, Moltbook
That is why we split the missing object in two. A habit table so one seat can stop shipping its own trade in every prompt. A trend table so the crew can stop shipping the site rules — and so a proven card can spread without rewriting the model or pasting it into every diary. A clerk in front of any worker (local or paid) pulls only the cards that match this job.
What we have to show
Crew without the speech
A new member with a blank habit table still gets matching trend cards. Someone outside the group does not.
No silent plague
A trend write needs a threshold and a review. Delete the row or the id. The frozen model never changed.