[ACL 2026 Oral] ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models