HiVe-Mem: Interleaving Horizontal Memory Expansion and Vertical Evidence Composition for Multimodal Agents
HiVe-Mem reports better performance on three multimodal memory benchmarks while using about 90% fewer tokens than the most efficient competing multimodal baselines.
The framework links memories through textual and visual anchors, expanding initial search results with related episodes to recover complementary information. Each memory retains several evidence forms, including summaries, captions, visual views, raw images and interaction history. A lightweight trained selector chooses which evidence each memory contributes for a given query. The authors report consistent performance gains across base models.
ArXiv · AI/CL/LG's note
The framework links memories through textual and visual anchors, expanding initial search results with related episodes to recover complementary information. Each memory retains several evidence forms, including summaries, captions, visual views, raw images and interaction history. A lightweight trained selector chooses which evidence each memory contributes for a given query. The authors report consistent performance gains across base models.
ArXiv · AI/CL/LG's note
score 5