memory-attack

This paper investigates memory manipulation in LLM-based agents for multiple-choice question answering, showing that corrupted memories can cause agents to select incorrect options even when the current query is clean.