HiVe-Mem: Interleaving Horizontal Memory Expansion and Vertical Evidence Composition for Multimodal Agents
The authors report better multimodal memory performance while using about 90% fewer tokens than the most efficient competing multimodal baselines.
HiVe-Mem links memory episodes through textual and visual anchors, letting retrieval reach related memories beyond its initial semantic matches. Each memory preserves several evidence forms, including summaries, captions, raw images and interaction history. A lightweight trained selector chooses which evidence each retrieved memory contributes for a given query. Across three multimodal memory benchmarks, the authors report consistent task improvements across base models.
HF Daily Papers' note
HiVe-Mem links memory episodes through textual and visual anchors, letting retrieval reach related memories beyond its initial semantic matches. Each memory preserves several evidence forms, including summaries, captions, raw images and interaction history. A lightweight trained selector chooses which evidence each retrieved memory contributes for a given query. Across three multimodal memory benchmarks, the authors report consistent task improvements across base models.
HF Daily Papers' note
score 4