Researchers have developed a new AI framework called Event-Causal RAG that improves the ability of large vision-language models to understand long videos. These models struggle with maintaining coherent event memory and recovering relationships in ultra-long videos. The proposed framework uses a dual visual-audio sentinel mechanism to segment video streams into semantically complete events, which are then stored in vector-graph memory. This allows for more accurate question answering on long videos. The researchers also introduce a new benchmark called ECV-1H that tests the ability of models to perform directed event-causal reasoning. According to the authors, their framework improves accuracy by 4.96%--11.67% across three open-source video foundation models. --- Why it matters: This matters because it addresses a significant challenge in AI research: understanding long videos with complex events and relationships. The proposed framework has the potential to improve the performance of various applications that rely on video analysis, such as surveillance, monitoring, and content creation. Source: https://arxiv.org/abs/2605.06185