WorldLines is a new benchmark for long-horizon stateful embodied agents. These agents must remember user routines and past interactions to assist humans over extended periods in real homes. The existing benchmarks mainly evaluate language-centric retrieval and question answering, but WorldLines focuses on embodied household assistance with dynamic environments. It constructs temporally extended household traces with dialogues, actions, and object state changes. A new memory framework called ObsMem is also proposed, which maintains visibility-aware memories for state-aware decisions. --- Why it matters: This matters to researchers in AI because it addresses the challenge of developing agents that can remember and adapt to long-term user routines in dynamic environments, a crucial aspect of household assistance. Source: https://arxiv.org/abs/2606.18847