Closed-Loop Memory Optimization Gives LLM Software Agents Cross-Task Learning and Lower Compute
Current LLM-based software engineering agents are episodic by design — they cannot retain or reuse experience across tasks, causing repeated context reconstruction and recurring mistakes. This paper introduces a closed-loop memory augmentation framework (MemOp) that grounds memory utility in validated downstream impact, serving simultaneously as a task-agnostic evaluation benchmark and an annotation-free optimization signal. Evaluated across both single-episode and cross-episode settings, the framework delivers measurable gains in success rate and resolve efficiency while meaningfully cutting computational cost.