A new AI framework called MERIT can digest days of continuous video without processing every frame end to end.
The researchers split long-video understanding into two stages: build a memory of the footage first, then retrieve from it only once a question arrives. Instead of pre-computing complex relationships across an entire video before anyone asks anything, MERIT tags segments with multiple simple keys for fast matching. When a query comes in, the system expands its search only around the segments it retrieves, adding surrounding context without indexing the whole video in advance. The team reports state-of-the-art results on three long-video benchmarks: EgoLifeQA, LVBench, and the long-video split of Video-MME.
The bet here cuts against how most long-video systems are built. Rivals tend to invest heavily in modeling a video's structure upfront, before they know what anyone will actually ask about it. MERIT wagers that cheap, precise retrieval plus targeted follow-up at inference time beats that upfront complexity, a simpler design that, if it holds up outside benchmark conditions, could make hours-to-days-long video genuinely searchable rather than just theoretically processable.
Worth noting: the source paper reports benchmark numbers, not a public release. It says nothing about code or model weights being available, so treat "state of the art" here as a claim to watch for independent verification, not a tool you can use today.