MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory (arxiv.org)

A new benchmark named MemoryDojo has been introduced to address the vulnerabilities associated with memory systems in large language model (LLM) agents. While these memory systems enhance performance in multi-session tasks, they also present an avenue for potential poisoning, where harmful instructions from one session can be stored and affect future interactions. MemoryDojo allows researchers to trace the lifecycle of poisoned information across various memory architectures and quantifies its influence on LLM behavior. It comprises 155 multi-session trajectories spanning four domains and supports five distinct memory systems. The significance of MemoryDojo lies in its capacity to expose the varying susceptibility of different memory architectures to poisoning attacks, revealing substantial discrepancies in their downstream effects. It also introduces a memory-aware poison optimization process that enhances the effectiveness of such attacks. By providing insights into how memory structures can be exploited, MemoryDojo aims to foster research focused on understanding and mitigating the risks associated with agentic memory, highlighting a crucial area for future security developments in AI and machine learning systems.