MemoryDojo: A Benchmark for Poisoning LLM Agent Memory Evaluation

Researchers introduce the MemoryDojo benchmark, which measures the impact of poisoned information on subsequent agent behavior by tracking its storage, propagation, and retrieval lifecycle in LLM agent memory. The benchmark contains 155 multi-session long-horizon trajectories covering four domains and supports five memory systems. The evaluation finds that all tested memory systems may be affected by memory poisoning, but downstream effects differ significantly. The authors also propose a memory-aware poisoning optimization method that can improve poisoning effectiveness, highlighting differences among memory architectures, and state that they hope the benchmark can support research on understanding and mitigating agent memory risks.