Researchers have developed StreamSoccer, a system for live soccer commentary that uses event-driven memory to process and generate commentary in real-time. The system integrates video streams with a fixed-budget active memory that retains completed event states and consolidates them into retrievable historical records. This allows the system to produce three types of commentary: current-event, recent-window, and historical-memory. StreamSoccer was tested on a dataset of 58 soccer matches and achieved high scores in all three commentary modes, with an average processing time of around 0.15 seconds per minute. --- Why it matters: This matters because it shows that event-driven memory can be effective for streaming video understanding tasks like live sports commentary, which requires models to process and generate content in real-time while managing growing amounts of data. Source: https://arxiv.org/abs/2608.19723