AdaMM Introduces Analytic Memory for Multimodal AI Agents

Long-term memory in multimodal AI agents traditionally focuses on retrieval, where interaction histories are summarized and indexed to return relevant information. While effective for recalling facts, this approach falls short when agents need to perform computations or derive insights from accumulated observations over time. This research introduces the concept of "analytic memory" as a complementary abstraction. Analytic memory organizes recurring multimodal observations into structured, queryable formats that support operations like filtering, aggregation, ranking, and temporal comparison. This allows agents to go beyond simple recall and perform more sophisticated data analysis. The AdaMM framework implements this by jointly supporting both retrieval and analytic memory. Instead of relying on predefined schemas, AdaMM extracts attribute-value observations from dialogues, images, and metadata, discovers recurring field structures, and materializes them for analytical access. A memory-aware planner then decomposes queries into appropriate retrieval or analytic operations, significantly improving performance on multimodal memory benchmarks.