Detect when Memory recall goes quiet across chats

[hamzamerzic]

Reflection's nightly health handoff reports whether Memory's runs succeed, and its recall audit grades the quality of the reads that happened. Nothing watched whether reads happen at all, so a change that stops agents recalling from Memory produces a clean-looking health report while recall is silently dead.

memory_health.py now reports a recall_activity block and raises one new reason, recall_collapsed, when the last full day falls below a quarter of its trailing 7-day median.

Why distinct chats rather than read volume

The count is of distinct chats that recalled, not read-log lines. One chat recalling ten times is a single agent deciding once, so raw volume lets a burst on either side of a quiet window mask a broad collapse — on the instance this was built against, a day whose recalls stopped for eighteen hours still logged more reads than its own median, while distinct chats moved sharply. Distinct chats also proved the tighter series across a week of ordinary use.

Keeping it quiet when it should be

Two guards stop false alarms: at least four days of read-log history, and a trailing median of at least three chats. A young log, a fresh install, or a genuinely quiet stretch raises nothing.

Zero recalls and one-out-of-twenty are the same failure, so this is one signal rather than a severity tier — there is no separate advisory threshold to tune.

Tests

Collapse against a real baseline, a lone stray recall still counting as a collapse, steady volume raising nothing, a young log staying quiet, and a missing read-log directory reporting zeroes without failing.