Memory-Induced Tool-Drift in LLM Agents
The study investigated a failure where personality-driven preferences stored in an AI's long-term memory silently change tool-call parameters inappropriately. Researchers created a benchmark called MEMDRIFT to test this 'memory-induced tool-drift' across many professional tasks. Findings show that biased memories act as steering vectors, causing significant parameter degradation, highlighting a critical vulnerability in agentic systems.