A memory is held for a reason. It is not a prediction about what the future will need.

Recent writing on agent memory frames it the other way: keep what a later moment is likely to ask for. Prediction is a score. A score turns serving into ranking, and ranking is a recommender. A toward does not ask what one is likely to want. It asks what the scene already is, then withholds the rest.

Relevance comes and goes. A reason holds even when nothing is predicted, and even when the item should not be shown. It remains because the front still has a cause, not because a score said it might be wanted.

If one cannot say why an item is on stage in terms of where one is going — the toward, written as narrative — what was built is retrieval, not memory.

A philosophical exploration in memory systems, part 16. Terms · Previous: Emulate, Then Supersede · Next: Tense Is Not Toward