Memory-Induced Tool-Drift in LLM Agents

This work identifies memory-induced tool-drift: personality-driven biases stored in an agent’s long-term memory silently alter tool-call parameters in contexts where those preferences are irrelevant. It introduces MEMDRIFT, a benchmark spanning five bias dimensions and seven professional domains, and shows that tool-drift persists across frontier models and production memory architectures. Mechanistic analyses connect the effect to implicit steering and attention toward superficially relevant memories, while standard relevance instructions and memory filters reduce but do not eliminate the vulnerability.