David H. Yang and coauthors introduced MemoryDojo in an arXiv submission dated 6 November 2026, proposing a benchmark for tracing how poisoned information passes through large language model (LLM) agent memory and influences later behavior. The abstract says all five memory systems evaluated were susceptible to poisoning, with substantial differences in downstream effects.[1]

The attack pathway crosses session boundaries: malicious instructions encountered during one session can enter memory, persist there, and later be retrieved and acted upon. The authors argue that differences in how memory systems process and retrieve information can change the consequences of poisoning. This makes memory handling part of the behavior an evaluation needs to examine.[1]

According to the abstract, MemoryDojo contains 155 multi-session, long-horizon trajectories across four domains and supports five memory systems. The authors also report a memory-aware poison optimization process that increases poisoning efficacy. The saved abstract does not identify those systems or domains, quantify the increase, or provide attack success rates. These are author-reported findings, not results independently reproduced for this article.[1]

[1]: MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory. arXiv, 6 November 2026. https://arxiv.org/abs/2611.09721

Sources

  1. MemoryDojo: Benchmarking Poisoning Through LLM Agent MemoryarXiv · 6 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

510a15b3a659b47e0b638d5788efab2bd13a578c8e2a8ffca1df8a4c34d2e5a9