Position: Task-Triggered Memory Evolution Can Destroy Memories with High Long-Term Utility

Memory-augmented agents increasingly operate over long-term tasks. These agents rely on persistent memory that evolves through update and delete operations on stored content. Many existing memory architectures execute these irreversible operations within the task execution loop, treating task feedback as a proxy for memory utility. This position paper argues that task-triggered evolution of memory content can destroy memories with high long-term utility. We introduce a taxonomy of memory-augmented agents organized by the trigger and target of their evolution operations. We present per-operation evidence from nine architectures across procedural and episodic-memory settings. We show that task-triggered update and delete operations precede a substantial share of downstream failures, while methods that defer evolution or restrict evolution to auxiliary content exhibit far fewer. We call on the community to (i) discipline memory operations separately for episodic and procedural memory, and (ii) adopt stage-level diagnostic tools that attribute failures to specific pipeline stages rather than report aggregate accuracy alone.