SWE-MeM boosts coding agents with adaptive memory management, hitting 60.2% resolve rate

SWE-MeM, introduced by researchers from multiple institutions, addresses a key bottleneck in long-horizon coding agents: managing lengthy and noisy interaction histories under limited context budgets. Existing methods rely on static compression or rigid constraints, failing to jointly optimize memory management and issue resolution. SWE-MeM provides a flexible memory tool that lets agents proactively decide when, what, and how to compress based on trajectory state, task progress, and remaining context budget.

Training uses synthesized proactive memory-management trajectories and a novel Memory-aware GRPO algorithm that jointly optimizes both memory management and issue resolution via trajectory splitting and step-level credit assignment. On SWE-Bench Verified, SWE-MeM achieves 43.4% resolve rate with a 4B model and 60.2% with a 30B model, outperforming all existing memory management baselines while reducing token usage. This dual improvement in performance and efficiency marks a significant step toward practical autonomous software engineering agents.