Semantic Drift

The compounding loss from summarising a summary. Each round is lossy against a base that already shrank, so the decay is geometric rather than linear.

Why it matters in a memory layer

It is the failure mode of the cheapest possible compaction loop — take yesterday's summary as today's input — and it is invisible, because the only record of what was lost was the input just replaced. Measured on this course's corpus, four naive rounds retain 19% of the original claims where re-deriving from anchors holds steady at 70%. The fix is not a better summariser; it is never summarising a summary.

Connections

Taught in: Semantic Drift

Used in: From Episode to Belief