New Benchmark Tests Multimodal Agents' Visual Memory.

While agent memory research has advanced significantly for text-based systems, the ability of multimodal agents to genuinely remember visual information in interactive environments remains largely unexplored. This paper addresses that gap by introducing DMV-Bench, an innovative interactive benchmark specifically designed to test the visual memory of long-horizon multimodal agents. The benchmark uses a controlled e-commerce catalog where product images contain unique, pre-rendered incidental cues, and agents are later tasked with recalling and navigating to specific cued products, ensuring the discriminative signal is purely visual. Inspired by dual-coding theory, the researchers also propose DualMem, a novel memory architecture that maintains both visual and verbal codes in parallel. Experiments on DMV-Bench demonstrate that DualMem consistently outperforms baseline captioning methods and three other recent multimodal agent-memory systems across various interaction lengths. This superior performance holds true for both Gemini 2.5 Flash and Qwen2.5-VL-7B models, even when controlling for memory-bank size and encoding-position bias. The findings highlight the critical role of visual memory in multimodal agents and offer a promising architectural approach to enhance it.