Streaming video understanding is difficult because a system must keep evidence before it knows what future questions will ask. Existing approaches often blur event transitions and drop relevant details when compressing video history, and they cannot recover visual evidence that was missed during observation.

DeltaStreamer tackles this with a two-stage design. Before a query arrives, it continuously watches the stream and explicitly records visual changes in structured memory, preserving how events unfold rather than only what is present. When a query arrives, a query-driven planner identifies the evidence needed and guides memory retrieval and targeted replay, allowing the agent to revisit historical frames and recover details omitted during observation.

The authors evaluated DeltaStreamer on the OVO-Bench and StreamingBench benchmarks, reporting overall scores of 66.07% and 59.53% respectively, with improvements over representative streaming video understanding methods. The paper is available on arXiv.