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

Researchers propose DeMem, a decision-centric memory framework that optimizes agent memory allocation based on preserving distinctions needed for sound decision-making rather than descriptive accuracy. Using rate-distortion theory, the approach identifies what information can be safely forgotten under memory constraints and demonstrates performance gains on long-horizon language agent tasks.

This research addresses a fundamental challenge in deploying language agents at scale: managing memory efficiently when computational budgets are fixed. Traditional memory systems prioritize descriptive metrics like relevance or salience, assuming that faithful representations of past events drive better decisions. The authors invert this assumption, arguing that memory's true value lies in maintaining distinctions between different histories that could lead to divergent decision outcomes. By framing memory management as a rate-distortion problem—a concept borrowed from information theory—the team establishes theoretical bounds on what information must be retained versus what can be safely discarded without degrading decision quality.

The practical contribution, DeMem, implements this theory through an online learning algorithm that dynamically refines its memory partitions only when new data indicate that merging states would create decision conflicts. This targeted approach contrasts with existing methods that continuously compress or summarize experiences, often losing critical distinctions. The framework comes with theoretical guarantees on regret, suggesting the approach scales reliably as agent complexity increases.

For the AI industry, this work has significant implications. Long-horizon agents deployed in real systems—from conversational AI to autonomous planning—operate under strict memory constraints due to latency and cost considerations. Current methods often sacrifice decision quality to stay within budgets. DeMem's decision-centric approach could enable agents to maintain performance while using less memory, directly reducing inference costs and improving response times. The research also provides a principled foundation for memory design, moving beyond heuristic compression schemes toward theoretically grounded optimization.

• →Memory should preserve decision-critical distinctions rather than accurately describing past events.
• →DeMem achieves consistent performance gains over existing methods under identical memory budgets on conversational benchmarks.
• →Rate-distortion theory provides an exact boundary for safe forgetting and optimal memory-decision quality tradeoffs.
• →The approach scales to long-horizon tasks where traditional memory mechanisms struggle with budget constraints.
• →Decision-centric memory design could reduce inference costs and improve latency for deployed language agents.