Metacognitive memory: mechanistic interpretability in the wild.

Most memory systems measure whether retrieval found the right content. We built a module that measures what the model did with it: what entered the workspace, what emotional geometry was active, what the probe reads outside the output pathway, and whether the retrieved content was actually processed at all. Four probes, integrated into a production agent, recording a geometric snapshot at every retrieval.

The pre-registered ordering — lived experience showing the strongest signal, then fictional, then scrambled, then nothing — appears in the data. At eleven memories, neither confirmatory test survives correction. We report the honest result and the parameters it hands the next, better-powered run.