WorldMemArena benchmark evaluates multimodal agent memory via action-world interaction
Researchers from UCSB, Stanford, ETH Zurich, Johns Hopkins, CMU, and J.P. Morgan Chase released WorldMemArena, a multimodal agent memory benchmark accepted at NeurIPS 2026. It evaluates memory through an observation-action-environment feedback loop, breaking long-term memory into write, maintain, retrieve, and use stages to pinpoint where agents fail. The benchmark includes multimodal cues, memory lifecycle tracking, and world changes after actions, addressing limitations of static recall or final QA evaluations.
Coverage timeline
机器之心机器之心
来自加州大学圣塔芭芭拉分校、斯坦福大学、苏黎世联邦理工学院、约翰斯·霍普金斯大学、卡内基梅隆大学和 J.P. Morgan Chase 的研究者，联合提出多模态 Agent 记忆评测基准 WorldMemArena。论文已被 NeurIPS 2026 接收；共同第一作者 Chengzhi Liu 与 Yuzhe Yang 均来自 UCSB。 团队盯住的是一个很实际的问题：Agent 明明保存了项目计划，预算更新后却仍按旧版本行动；看过环境截图，下一步又忽略了画面里已改变的状态。它究竟是没记下来、没及时更新，还是找到了证据却没用上？ 能答出“以前说过什么”，不等于能用好记忆 。这正是 WorldMemArena 想追问的核心问题。 论文： WorldMemArena: Evaluating Multimodal Agent Memory Through Action–World Interaction 论文地址：https://arxiv.org/abs/2605.29341 项目主页： https://worldmemarena-mem.github.io/ 代码：https://github.com/UCSB-AI/WorldMemArena 数据：https://huggingface.co/datasets/LCZZZZ/WorldMemArena 只看最后答对没有 漏掉了哪一步？ 很多记忆评测把长对话交给模型，最后用问答正确率打分。但一次答错，可能是当初没写下关键事实，也可能是新信息来了却没有覆盖旧信息；还可能是证据明明存着，检索时没找出来，或找出来后没有用进回答。把图像先压成文字描述，也容易丢掉位置、时间和操作过程中的细节。 WorldMemArena 将这些问题放到“ 观察—行动—环境反馈 ”的循环里：Agent 不只回忆对话，还要面对会变化的任务状态、网页、文件、截图与工具结果。评测因此不再只看终点，而是追踪记忆怎样影响后续判断。 图 1｜传统评测常聚焦静态回忆或最终问答；WorldMemArena 加入多模态线索、记忆生命周期与行动后的世界变化。图源：论文 Figure 2。 把“记忆”拆成四个可以定位的环节 论文提出 Action–World Interaction Loop ，把长期记忆拆成四步： 写入 ，从当前观察中挑出以后可能有用的证据； 维
Related stories
Fudan, Tencent Hunyuan, Tsinghua Release ExplorationBench to Test AI Rule Discovery
Fudan University, Tencent Hunyuan, and Tsinghua University released ExplorationBench, a benchmark measuring how AI systems discover rules in unfamiliar environments. It includes two altered 'alien worlds'—the programming language AlienCode and the logic proof system AlienLogic—where models must learn rules through trial and error. Results showed that without environment feedback, scores remained low (max 11.0%), while autonomous exploration led to high scores (up to 89.0%).
2 reports · 1 sources
STILL: Content-Aware Intra-Layer Hybrid Attention Cuts Memory 80%, Speeds Prefill 83%
Researchers from Harbin Institute of Technology (Shenzhen) and collaborators present STILL, a NeurIPS 2026-accepted architecture that routes tokens to Softmax or linear attention within the same layer based on content importance via a self-saliency score. It achieves 80% memory reduction and 83% prefilling speedup at 16K context, with only 0.04B training tokens for linearization, and up to 86.2% improvement over prior linearization methods on long-context benchmarks.
MemTensor Releases MemT 1.0, First Memory-Native Foundation Model with Open-Source Checkpoints
On October 20, 2026, MemTensor released MemT 1.0, the first memory-native foundation model, which encodes historical information into a continuously updatable internal memory state that participates in subsequent generation. The release includes MemT-1.0-Flash (288B parameters, 17B active, 336M dynamic memory) and MemT-1.0-Nano (36B parameters, 3B active, 10M dynamic memory), described as the largest open-source parametric memory models to date, with checkpoints and a technical report available on GitHub, HuggingFace, and ModelScope.
Meta launches Muse Wonder and Muse Wonder Full AI models
Meta has released two new AI models, Muse Wonder and Muse Wonder Full, claiming state-of-the-art performance on personal and professional agentic tasks. The models introduce Active Memory and Active Focus features and are available to Muse Code subscribers, with Muse Wonder also accessible via the Meta Model API. Meta plans to eventually power its Muse AI agent with Wonder, starting with subscribers.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe