WorldMemArena: Evaluating Multimodal Agent Memory Through Action-World Interaction

Researchers introduced WorldMemArena, a benchmark with 400 multimodal multi-session tasks for testing agent memory in an action-world interaction loop.

It adds stage-level labels for memory points, updates, distractors, and evidence chains to better evaluate how agents remember and use information.

Results comparing long-context models, hand-built memory systems, and harness-based memory agents show that stronger memory writing does not always improve performance.

The study finds current systems still struggle to use visual evidence and to stay stable across changing domains.