Deduplication

Collapsing records that say the same thing. Not records that are about the same thing — that is entity resolution — and not the same turn processed twice, which is idempotency.

Why it matters in a memory layer

The three are easy to conflate and fail differently. Idempotency is settled at write time by a content-addressed id and catches re-ingestion of one turn. Deduplication catches the same fact arriving from different turns, where the sources differ, the ids differ, and idempotency is blind by construction.

The cost of the two mistakes is not symmetric. A missed duplicate wastes a slot in the token budget. A wrong merge destroys a distinction — and the tempting near-misses are refinements and contradictions, which score high precisely because they are about the same subject. So the threshold belongs near certainty, and everything below it is a relationship for conflict detection to name rather than a repetition to collapse.