Why AI Agent Memory Systems Are Structurally Trapped in Feedback Loops

A technical analysis argues that recall in agent memory systems is not a passive read operation but a write operation, because retrieving a memory updates its usage score, timestamp, or reinforcement weight. This creates a closed positive feedback loop where frequently retrieved notes become increasingly dominant in future rankings, regardless of their accuracy. The structural problem is that the same mechanism meant to surface relevant memories also prevents the system from demoting incorrect or outdated ones. Usage-based ranking is acknowledged as the best available signal for memory relevance, since alternatives like age, explicit labels, or content scoring are all unreliable proxies. The core argument is not that usage-weighting should be removed, but that its feedback dynamics are rarely examined or designed against deliberately.