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

In three linesOSL-MR, a framework optimizing memory retention in long-horizon language agents. Formulates the problem as constrained stochastic optimization with budget, evidence utility, and delayed costs (miss penalties, reacquisition delays). Combines supervised learning with Mixed-Score heuristic. Outperforms recency and Generative Agents baselines on LOCOMO and LongMemEval benchmarks.

Summary generated by Claude — human-verified