Field note · 2026-09-10
When the prompt files fight each other
An agent's standing files compete with each other and with fresh memory for the same space. In mine, three of four people got no fresh facts, and nothing logged an error.
- 3 of 4
- people who received no fresh facts because pinned memory filled the budget2026-09-10
- 0
- errors logged while that happened2026-09-10
- 31,937 → 9,811
- characters in the main instruction file after a diet2026-08-26
- 0.75 days
- for a 1,400-token prompt saving to repay the cost of changing it2026-08-13
What I saw
Bob knew old things about people and not new ones. He also sounded, one crew member said, like mandatory training: numbered coping lists and worksheet headings.
What caused it
Two different failures in the same files. First, pinned memory was exempt from the packet budget, and it grew. The pinned heads for four people measured between 13,172 and 20,817 characters against a 14,000-character budget, so three of the four received no unpinned facts at all. The builder logged zero errors.
Second, the voice. I ruled out a loading failure: the persona files were present and whole in the offending turns. One class of turn, advice about another person, pushed the model into a counselling register, and no guidance covered it.
What I changed
Seven bootstrap files load in a fixed order, with a cap per file and in total. I put the main instruction file on a diet, from 31,937 characters to 9,811, and set a rule for anything that loads on every turn: will this sentence still be true in a year? If I'm unsure, it goes in the person's file.
Promotion into a standing file is refused without a source, a duration, and a speaker. Every release declares how the agent learns of the change, and anything always-on needs evidence of its prompt cost. Prompt edits are costed too: rewriting an instruction costs about 200,000 uncached tokens once, so a 1,400-token saving repays in under a day and a 100-token saving takes ten.
For the voice, one rule ("advice is not a worksheet") and a deterministic detector that triggers a single rewrite, with a check that names, numbers, and dates survive.
What I'd tell another team
Never put an example answer in guidance. Bob repeated mine word for word. And be suspicious of a builder that reports no errors; mine was correct, and three people were getting nothing.