In recent years, memory has become a core component of long-horizon AI agents, enabling them to reuse past experiences when operating web browsers, software tools, and other interactive environments. However, existing work primarily treats memory as a supply issue, focusing on what experiences to write, how to store them, and which entry to retrieve for the next task. We still lack a clear account of how models consume retrieved memory across a multi-step action trajectory. This consumption process is crucial, as it determines not only which memories should be retrieved but also what models and control policies are necessary for their safe usage.
To diagnose this process, we propose the Entry-Propagation-Recovery (E-P-R) framework, which examines where memory first alters an action, whether that change persists, and whether the agent can recover after deviating from the correct path. We instantiate E-P-R on WebArena and on MemTrapBench, a controlled benchmark we created to isolate these phases. Our findings indicate that failures often begin at entry: agents adopt conflicting memory at the first exposed decision point, even when it is task-incorrect. Repeated exposure amplifies this early error, while recovery after divergence is weak. Collectively, these effects create a compliance trap: across models, conflicting memory induces similar compliance rates, but once agents comply, their success rates collapse to a low floor. Consequently, stronger agents suffer larger absolute damage because each compliance event erases more baseline capability. These results suggest that memory-augmented agents should be evaluated not only by retrieval quality or final success rate but also by how they consume memory throughout the trajectory.
Blogger's Review: This paper delves into the complexities of how AI agents handle memory, especially regarding decision-making in the face of conflicting information. The E-P-R framework offers a robust tool for analyzing this phenomenon, underscoring the importance of memory management in executing complex tasks. Future research should further explore optimizing memory consumption processes to enhance the overall performance of AI systems.