Researchers have created a benchmark called EgoMemReason to evaluate the ability of artificial intelligence models to reason over long periods of time using memory. The benchmark focuses on egocentric video understanding, where models must integrate information from hours or days of continuous visual experience. It tests three types of memory: entity memory, event memory, and behavior memory, and has been evaluated on 17 different methods, which achieved an average accuracy of only 39.6%. This suggests that long-horizon memory remains a challenging problem in AI research. --- Why it matters: This matters to researchers in AI because it highlights the difficulty of developing models that can reason over extended periods of time using memory. It has implications for applications such as smart glasses, embodied agents, and life-logging systems that require integrating information from long video sequences. Source: https://arxiv.org/abs/2605.09874