Beyond Retrieval: Analytic Memory for Multimodal Agents
AdaMM adds structured analysis to long-term multimodal memory, not just lookup.
The paper defines “analytic memory” as memory that can filter, aggregate, rank, and compare observations over time. AdaMM extracts provenance-linked attribute-value observations from dialogue, images, and metadata, then materializes recurring fields for analytical queries. A planner routes each query step to either retrieval or analytic tools. On MemEye and MemGallery, the authors report gains of up to 11.3% and 7.3%. ArXiv · AI/CL/LG's note
The paper defines “analytic memory” as memory that can filter, aggregate, rank, and compare observations over time. AdaMM extracts provenance-linked attribute-value observations from dialogue, images, and metadata, then materializes recurring fields for analytical queries. A planner routes each query step to either retrieval or analytic tools. On MemEye and MemGallery, the authors report gains of up to 11.3% and 7.3%. ArXiv · AI/CL/LG's note
score 5