Control-Plane Placement Shapes Forgetting: An Architectural Study of Agent Memory Across Thirteen System Configurations

The study presents findings on the impact of control-plane placement in agent memory architectures, analyzing thirteen configurations that influence forgetting failure modes in LLMs. Key results indicate that mutation-time hooks significantly enhance intent-aware deletion (78-85%) and improve overall performance across categories (91.7-93.2%) compared to deterministic approaches. The introduction of ForgetEval, a comprehensive benchmarking suite, highlights the predominance of forgetting failures in production systems and offers new metrics for evaluating memory performance, which is crucial for practitioners focused on optimizing LLM integration in memory-intensive applications.