When to Forget: A Memory Governance Primitive

Researchers propose Memory Worth (MW), a dynamic metric for managing agent memory systems. MW tracks how often memories correlate with successful versus failed outcomes, providing a lightweight basis for detecting outdated information and optimizing memory retrieval. The metric converges to conditional success probability and achieves 0.89 correlation with actual memory usefulness in tests, significantly outperforming static systems. This approach requires only two counters per memory unit.