arxiv:2609.28236

EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks

Published on Sep 23

· Submitted by

Lizhou Liang on Sep 29

ZJU-OmniAI

Upvote

Authors:

Lizhou Liang ,

Abstract

Researchers introduce a benchmark and external memory system to evaluate and improve long-horizon embodied agents' ability to retain and update environmental information during interaction.

Generated by thinkingmachines/Inkling-Small

Long-horizon embodied interaction requires agents to retain and continually update information about the environment as they observe, act, and encounter change. Yet current agents struggle to maintain such memory reliably. Our analysis traces this limitation to four key deficiencies: weak fine-grained visual memory, unreliable dynamic world-state tracking, failing to record world state revealed by interaction outcomes, and limited generalization from prior experience. However, existing benchmarks do not directly assess these memory capabilities during long-horizon embodied interaction. To address this gap, we introduce EmbodiedMemory-Bench (EMem-Bench), comprising 2,554 interactive episodes across four task families. EMem-Bench requires agents to build and update memory from interaction history, then use it to complete a later task by acting in the environment. We further present Embodied-Memorizer (EMem), an external memory system that organizes embodied experience into spatial, event, and scene memories. We also train EMem-8B, an 8B policy that manages and uses these memories. We evaluate a diverse range of open-source and proprietary MLLMs and representative multimodal memory systems. Results show that current models remain weak and uneven across the four challenges. Under matched backbones, EMem achieves the best overall performance among the evaluated memory systems and improves both open-source and proprietary models, while EMem-8B further improves over its backbone. Project page: https://zju-omniai.github.io/EmbodiedMemoryBench/

View arXiv page View PDF Project page Add to collection

Community

lzLiang

Paper author Paper submitter Sep 29

Excited to share EmbodiedMemory-Bench! 🚀

We study embodied memory for long-horizon interactive tasks, where agents must not only remember past observations, but also continuously update world states, learn from interaction outcomes, and reuse experience for future decisions. EMem-Bench introduces 2,554 executable episodes across four memory challenges, together with Embodied-Memorizer, a structured memory framework combining spatial, scene, and event memory.

We hope this work helps move embodied agents from simply “seeing more history” toward truly maintaining and using a persistent, evolving memory of the world.

Project, code, models, and dataset are all open — feedback and discussions are very welcome!

librarian-bot

Sep 30

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

Jethro37

22 days ago

Thank you for sharing EmbodiedMemory-Bench. Separating recall and state updating from failure feedback and experience generalization helps distinguish remembering an episode from transferring a correction to a new situation. This informs our discussion of memory evaluation in the working revision of In-Context Learning for Robots: Methods and Applications (https://arxiv.org/abs/2609.36012), drawing on the benchmark's executable simulation setting. We hope the survey offers useful context for your work. If you find it helpful, we would appreciate your considering it as a reference. Thanks again!

zwq2018

15 days ago

Thank you for sharing EmbodiedMemory-Bench. Separating recall and state updating from failure feedback and experience generalization helps distinguish remembering an episode from transferring a correction to a new situation. This informs our discussion of memory evaluation in the working revision of In-Context Learning for Robots: Methods and Applications (https://arxiv.org/abs/2609.36012), drawing on the benchmark's executable simulation setting. We hope the survey offers useful context for your work. If you find it helpful, we would appreciate your considering it as a reference. Thanks again!

Thank you for the recommendation. We will read your paper carefully and will update our paper accordingly in a future version.