Researchers from UC Santa Barbara, Stanford, ETH Zurich, Johns Hopkins, Carnegie Mellon and J.P. Morgan Chase have jointly proposed WorldMemArena, a multimodal agent memory benchmark, according to a 机器之心 report. The paper, whose co-first authors Chengzhi Liu and Yuzhe Yang are both at UCSB, was accepted at NeurIPS 2026. WorldMemArena evaluates memory inside an observe–act–environment-feedback loop and splits long-term memory into writing, maintenance, retrieval and use stages, with more than 400 multi-session tasks, 24,258 QA pairs and 15,595 images or screenshots. In the paper's results, M2A reached 86.83% memory recall but only 50.14% final QA accuracy, while MemGPT scored 85.20% and 57.81%, and systems tended to append new memories rather than revise outdated ones and did not reliably use visual information in cross-modal reasoning.