A new benchmark called ICM-Bench, described in a recent arXiv paper, focuses on a specific gap in AI agent evaluation: remembering who participated in past events, not just what happened. The authors argue that long-horizon multimodal agents need to track people across time and modalities.

According to the abstract, this capability depends on linking recurring faces, voices, names, person-associated objects, events, and social relations. ICM-Bench appears designed to test these connections in a long-term memory setting, though the provided abstract does not give details on specific tasks or results.

The work highlights a shift from testing factual recall to testing identity reasoning in agents that operate over long horizons. By focusing on person-level memory, the benchmark could help developers build agents that maintain coherent social awareness.