[ACL 2026 Best Resource Paper] ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models