Give agent memory three states, active, dormant and retired, with asymmetric thresholds so entries stop oscillating

arXiv 2608.18177·medium signalXBlueskyLinkedInCopy link

This paper argues that continual learning's bias toward retention is wrong for enterprise agents, where policies, tools, and market conditions expire and then sometimes become relevant again. The proposed Hysteretic Reversible Memory Controller accumulates relevance signals and uses different thresholds for suppressing and reactivating an entry, which is what prevents a memory from flapping between states, and it tests reactivation in a non-production environment before promotion. The trading example makes the case, a strategy that stops working in one market cycle should go dormant rather than be deleted.