Learning What to Remember: Observability-Safe Memory Retention via Constrained Optimization for Long-Horizon Language Agents

The paper introduces OSL-MR (Observability-Safe Learning for Memory Retention), a framework designed to optimize memory retention for long-horizon language agents by modeling it as a constrained stochastic optimization problem. OSL-MR employs a Mixed-Score heuristic and an evidence learner to balance online observability and offline supervision, significantly outperforming recency-based methods and Generative Agents-style scoring in benchmarks like LOCOMO and LongMemEval, particularly under strict memory constraints. This approach is crucial for practitioners as it enhances memory management strategies in AI systems, ensuring efficient resource allocation while maintaining performance in real-world applications.