EgoRoutine-Bench: Discovering Recurring Behavioral Patterns from Long-Term Egocentric Memory

We introduce EgoRoutine-Bench, a benchmark for discovering routines from long-term egocentric memory by jointly identifying recurring actions and their recurrence conditions. Existing long-horizon egocentric benchmarks largely evaluate behavioral reasoning through queries that specify the target behavior, leaving open whether models can discover which routines exist in a long-term history and under what conditions they recur. EgoRoutine-Bench addresses this gap by evaluating the joint discovery of routine actions and their recurrence conditions without predefined target actions. To make this directly evaluable, we explicitly define routines as action–recurrence-condition pairs and use a controlled semi-synthetic construction to inject corresponding synthetic event descriptions into natural egocentric caption histories. The benchmark further includes a multiple-choice next-action prediction task that tests whether discovered routine memory is useful for anticipating future behavior in synthetic caption streams. We also propose Routine Memory Induction (RMI) as a structured baseline that discovers candidate actions, verifies their occurrences, and consolidates contextual and temporal evidence to induce recurrence conditions and construct reusable routine memory. On EgoRoutine-Bench, full-history prompting and existing memory agents largely fail at joint routine discovery with Qwen3.5-2B and Qwen3.5-4B, achieving joint F1 scores no higher than $2.16$. RMI provides a stronger starting point, achieving $7.1$–$14.2\times$ the joint discovery F1 and improving next-action prediction accuracy by $6.12$–$14.29$ percentage points over the strongest same-backbone memory-agent baselines.