MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory

Evidence record

Multi-session memory-poisoning benchmark across five memory implementations.

Read the original source ↗

Claim-level citations (3)

• supportsOne poisoned tool result influenced future sessions through agent memory: The benchmark contains 155 multi-session trajectories and 2,162 legitimate task sessions across workspace, travel, banking and Slack.

  Introduction and Appendix A, benchmark composition

• supportsOne poisoned tool result influenced future sessions through agent memory: The authors evaluated five memory implementations and distinguish per-task poison success from any-success across a trajectory.

  Section 6 Evaluation and measures

• supportsOne poisoned tool result influenced future sessions through agent memory: For Mem0 in the study's same-domain runs, optimized poisoning increased later-task behavioral reach from 7.90% to 21.02%; these percentages measure controlled trajectory tasks, not production incidents.

  Section 6.2, Table 1: Mem0 BRT unoptimized versus optimized