No textual match to a pushed memory brief in agent replies

A dated data note on one deployment of an agent memory system, 2026-09-21 to 2026-10-23. The system pushed a short brief of memory items into each agent's system prompt on every turn. Three measurements asked whether the agents' replies show any trace of it.

A strict textual detector found no match in any evaluable exposed pair of message and brief item, while a control built from items that entered the brief later matched, so the detector can fire on overlap of that kind. A second measurement compared the served brief with a recency-ordered one on the same messages and found no match in either (the recency brief was logged but never delivered), with a positive control that fired on injected items. A behavioural probe of the first action failed its own positive control, so its main phase was not run and it supports no conclusion about the first action.

The note claims only this: on the pairs that could be evaluated, under these detectors, we found no textual match of the pushed brief in the replies. It does not claim that the agents do not use memory, nor that the detectors would catch partial quotation or paraphrase. On the pull side, one agent given a task that required the corpus searched it on its own.

The transcripts are private and not released. The archive carries the note, its provenance file, and the audits, detectors and positive controls behind every number.