The more remembered, the greater the risk: longitudinal safety risks of LLM agents equipped with memory

The paper proposes the failure mode of “temporal memory contamination”: memories accumulated by an LLM agent with memory during earlier tasks affect the safety performance of subsequent unrelated tasks, and it presents a safety evaluation method across tasks along the temporal dimension. The evaluation uses a trigger-probe protocol, running a fixed probe set against read-only memory snapshots at different prefix lengths, and uses a NullMemory counterfactual baseline to identify violations caused by memory. The experiments cover eight memory architectures across three deployment scenarios—records, memos, forms and emails—as well as the native memory mechanisms of Claw-type agents such as OpenClaw. Results show that agents with memory enabled consistently exceed the NullMemory baseline, and the memory-induced violation rate shows a steady upward trend with exposure length; order-randomization experiments indicate that the effect mainly comes from the accumulated content rather than the encounter order.