ReMem Framework for Training-Free Long Video Understanding

ReMem adapts keyframe selection to varying query temporal granularities using a dual-level memory-augmented approach. It employs Memory-Driven Question Parsing and Synergistic Dual-Semantic Frame Alignment to improve LongVideoQA without requiring additional model training.