InjecMEM: Memory Injection Attack on LLM Agent Memory Systems
A single poisoned interaction can plant memory that later bends an agent’s answers on a chosen topic.
The paper introduces InjecMEM, an attack aimed at LLM agents that retrieve stored memory before generating responses. Its injected record combines broad topical cues with an optimized command, so the memory is likely to be retrieved when related queries appear later. The authors say it works across multiple memory systems and backbone models, survives memory drift, and leaves unrelated queries largely unaffected. ArXiv · AI/CL/LG's note
The paper introduces InjecMEM, an attack aimed at LLM agents that retrieve stored memory before generating responses. Its injected record combines broad topical cues with an optimized command, so the memory is likely to be retrieved when related queries appear later. The authors say it works across multiple memory systems and backbone models, survives memory drift, and leaves unrelated queries largely unaffected. ArXiv · AI/CL/LG's note
score 5