Long videos pose a challenge for multimodal large language models (MLLMs) because they often exceed the models' context capacity. Agentic search systems that answer queries about such videos typically repeat costly video processing and grounding for each new query, making the process inefficient.
The paper proposes a "write once, query many" approach: process the video once and store a compact, evolving reusable memory. This memory can then serve many queries, amortizing the initial processing cost across all of them.
The abstract notes that reusable memory must preserve certain information, though the description cuts off there. The full paper presumably details how the memory is constructed and updated to maintain the needed information over time.