Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory

Researchers have introduced a novel framework for long-horizon language agent memory, challenging existing mechanisms that prioritize descriptive criteria like relevance or summary quality. The study argues that memory value lies in preserving distinctions between histories necessary for optimal decision-making under limited runtime budgets. By framing this as a decision-centric rate-distortion problem, the authors measure memory quality by the loss in decision quality caused by compression. This approach defines an exact forgetting boundary and a memory-distortion frontier to balance memory budget and performance. The team proposes DeMem, an online memory learner that refines data partitions only when shared states induce decision conflict, offering near-minimax regret guarantees. Empirical tests on synthetic diagnostics and conversational benchmarks demonstrate that DeMem achieves consistent performance gains within fixed runtime constraints. The findings support the principle that agent memory should prioritize decision-critical distinctions over faithful historical descriptions, marking a significant advancement in artificial intelligence memory management.