A Memory System That Got Slow, Expensive, and Worse

A product team added "the agent remembers everything you tell it" as a feature. Six months later: every user's memory profile has grown to hundreds of stored facts, all of which are loaded into every single prompt because "we don't want to risk missing something the user said." Latency and cost per request have risen sharply, and users report the agent now seems to ignore or contradict things they told it recently — even though those facts are, technically, in the loaded memory.

1. Diagnose both problems — cost/latency and the "seems to ignore recent facts" complaint — in terms of the eager-vs-retrieval- triggered trade-off.
2. Redesign the loading strategy: what stays eager, what becomes retrieval-triggered, and what criteria decide which bucket a given fact falls into?
3. Separately, propose a fix for facts that were true when stored but may no longer be — is that the same problem as (1) and (2), or a different one?