WorldLines: Benchmarking and Modeling Long-Horizon Stateful Embodied Agents

WorldLines, from HKUST(GZ) and Knowin, targets a gap in embodied AI evaluation: long-horizon household assistance where agents must preserve routines, object locations, device states, and partial observability across days rather than solve isolated short episodes. The benchmark builds evidence-linked samples from Habitat/HSSD household scenes using a project-driven pipeline that generates cross-day traces with dialogue, human activity, robot actions, execution feedback, and state transitions, then converts them into Memory QA and Embodied Task Planning tasks. To address the mismatch between flat text memory and stateful embodied reasoning, the paper proposes ObsMem, which separates event evidence, structured world-state trails, belief status, and commitments, and routes each query through typed retrieval plus evidence selection. On 310 Memory QA samples, ObsMem reaches a Judge score of 0.713 and Perfect Rate of 69%, outperforming A-mem at 0.575 and 53% and improving Event R@5 to 0.537 versus 0.378 for Mem0; on the hardest StateMH-E setting it climbs to 0.452 versus 0.264 for Mem0. In a 21-sample planning probe, ObsMem also leads with a Plan Judge of 0.684, showing that observer-grounded memory transfers better into executable household plans than Mem0, MemoryOS, or graph-based memory.