DeMem: A Decision-Centric Memory Framework for Language Agents

A recent study published on arXiv (2605.10870) presents a decision-focused rate-distortion framework for agent memory. The researchers contend that current memory systems emphasize descriptive factors such as relevance and salience; however, the core utility of memory is in maintaining separations between histories that need to be distinct within a constrained budget to facilitate sound decision-making. They define this challenge as a rate-distortion issue, quantifying memory effectiveness by the degradation in decision quality resulting from compression. This leads to a precise forgetting threshold for what can be omitted and a memory-distortion boundary that illustrates the best balance between memory resources and decision quality. They introduce DeMem, an online memory learner that adjusts its partitions only when data indicates that a shared state could lead to decision errors. This research is aimed at long-horizon language agents with limited runtime memory.