Long-horizon embodied tasks require agents to hold on to what they have seen, keep updating their picture of the world as things change, and put earlier experience to use later. According to a new research paper, today's agents struggle to do this reliably.
The authors trace the problem to four specific deficiencies: weak fine-grained visual memory, unreliable tracking of a changing world state, failure to record world state revealed by the outcomes of actions, and limited generalization from prior experience. Existing benchmarks, they argue, do not directly measure these abilities during long-horizon embodied interaction.
To close that gap, the team introduces EmbodiedMemory-Bench (EMem-Bench), a benchmark of 2,554 interactive episodes spanning four task families. Agents must build and update memory from their interaction history, then act in the environment to complete a later task.
The paper also presents Embodied-Memorizer (EMem), an external memory system that organizes embodied experience into spatial, event, and scene memories, plus EMem-8B, an 8B policy trained to manage and use those memories.
Evaluations across a range of open-source and proprietary multimodal LLMs and representative multimodal memory systems show current models remain weak and uneven across the four challenges. Under matched backbones, EMem delivered the best overall performance among the evaluated memory systems, improving both open-source and proprietary models, while EMem-8B improved further over its backbone.
The project page, code, models, and dataset are released openly, and the authors frame the work as a step from agents that merely "see more history" toward those that maintain and use a persistent, evolving memory of the world.