event-centric

EM^2Mem proposes an event-centric multimodal memory framework that binds heterogeneous evidence to event anchors for compact, generation-ready memory in long-video question answering, improving accuracy and reducing latency.