HABIT-Bench: Evaluating Habit Induction from Longitudinal Weak Evidence in Agent Memory.

arXiv:2610.24520v2 Announce Type: replace Abstract: In long-term interactions with large language model (LLM) agents, users reveal habits through their choices and actions. These habits capture implicit user information that agents must infer across conversati