PyraVid uses brain-inspired pyramid memory for long video AI reasoning

Researchers from multiple institutions have introduced PyraVid, a hierarchical multimodal memory framework designed to address the challenge of long-horizon video reasoning in agentic systems. While prior work focused on unimodal memory, PyraVid tackles the complexities of integrating heterogeneous inputs—such as video, audio, and text—while aligning person-centric information. Inspired by Event Segmentation Theory from cognitive science, the framework organizes long videos into a coarse-to-fine pyramid structure, allowing agents to access memories at different granularities and aggregate evidence effectively.

PyraVid also introduces structure-guided memory expansion with pruning, which retrieves causally connected events even when semantic similarity is low—reducing noise and improving recall. In experiments across multiple long-video benchmarks, PyraVid consistently outperformed baseline methods across model scales and question types. This work represents a significant step toward enabling AI agents to reason over hours of real-world video data, with applications in autonomous systems, surveillance, and media analysis.