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

Researchers introduce OSL-MR, a framework that optimizes memory retention for long-horizon language agents by treating it as a constrained optimization problem rather than local decisions. The approach combines learned evidence valuation with heuristic scoring while respecting real-world observability constraints, demonstrating superior performance over existing methods on benchmark datasets.

This research addresses a fundamental challenge in deploying language agents at scale: managing finite context windows when agents accumulate extensive observations, reasoning traces, and retrieved information over extended task horizons. The problem has grown increasingly critical as AI systems tackle more complex, multi-step reasoning tasks that require maintaining coherent memory of past interactions. OSL-MR's contribution lies in moving beyond reactive, heuristic-based memory management toward a principled optimization framework that explicitly models long-term consequences of retention decisions.

Previous approaches relied on simple scoring mechanisms or learned compression without systematically accounting for the costs of forgetting information—including reacquisition delays, penalties from missed context, and risks from stale data. The separation between online-observable features and offline supervision is particularly noteworthy, as it enables practical deployment without requiring information unavailable during inference. This distinction reflects real-world constraints that many AI systems face when operating with limited observability.

The framework's performance gains become pronounced under tight memory budgets, a scenario increasingly relevant for edge deployments and cost-conscious applications. The Mixed-Score heuristic serving dual purposes—as a deployable baseline and an inductive prior for learning—demonstrates thoughtful engineering that balances theory with practical constraints. For developers building agent systems, this research suggests that memory management deserves explicit optimization rather than ad-hoc solutions. The benchmark improvements validate that principled approaches outperform intuitive baselines across varying cost configurations.

• →OSL-MR formulates memory retention as a constrained stochastic optimization problem with explicit budget, utility, and delayed-cost modeling.
• →The framework enforces strict separation between online-observable and offline-available features, enabling safe deployment under realistic constraints.
• →Experiments demonstrate consistent performance improvements over recency-based and heuristic baselines, particularly when memory budgets are constrained.
• →Mixed-Score heuristic improves precision while maintaining recall without requiring additional observability during inference.
• →Sensitivity analysis shows robustness across diverse cost configurations, suggesting broad applicability to different deployment scenarios.