ViSAGE Improves Long-Form Video Understanding with Self-Correcting Memories

A new research paper introduces ViSAGE, a novel multimodal agentic memory framework designed to enhance long-form video understanding. Existing AI agents often struggle with maintaining consistent entity identities and temporal grounding over extended video sequences, frequently losing fine-grained cues due to aggressive data compression or relying on vector similarity that can lead to confusion. ViSAGE addresses these challenges by constructing self-correcting, entity-centric memories. It achieves this by anchoring entity identity through cross-modal binding across long temporal ranges within a video. The framework then employs bidirectional memory refinement, allowing delayed identity evidence to retroactively update historical records and improve future reasoning accuracy. Furthermore, ViSAGE incorporates multi-agent cross-verification. This mechanism assesses retrieved evidence against an identity-evidence alignment constraint, enabling the system to abstain from providing unsupported answers when evidence is insufficient, thereby reducing hallucination. Extensive evaluations demonstrate that ViSAGE significantly outperforms current state-of-the-art baselines, achieving a 5.9% higher accuracy.