ViSAGE framework cuts video AI errors by 5.9% with self-correcting memory

ViSAGE is a new multimodal agentic memory framework developed by Xinkui Zhao and six co-authors, designed to address a critical flaw in long-horizon video understanding: entity confusion. Current agentic memory systems rely on aggressive compression and segment-wise processing, which discards fine-grained identity cues, and they use vector similarity retrieval that often surfaces semantically related but identity-mismatched evidence. This leads to error propagation and hallucinated answers in tasks requiring entity-consistent, temporally grounded reasoning.

ViSAGE tackles this with three key innovations. First, it anchors entity identity via cross-modal binding over long temporal ranges, linking visual and textual cues to a stable identity. Second, it applies bidirectional memory refinement, which propagates delayed identity evidence backward to retroactively unify historical records and improves future reasoning. Third, it introduces multi-agent cross-verification that assesses retrieved evidence under an identity-evidence alignment constraint, allowing the system to abstain from answering when evidence is missing rather than hallucinate. In extensive tests, ViSAGE consistently outperformed the strongest baseline, achieving 5.9% higher accuracy. The paper has been accepted at ACMMM 2026, marking a significant step toward reliable long-form video understanding for AI agents.