JointCore preprint reports fewer queries to select AI agent memories

Researchers report in an arXiv preprint that JointCore, a method for selecting which memory records an AI agent needs, reduced model queries while preserving full-memory behavior in benchmark tests. Across four model families with 12B-70B parameters and 432 tasks per model, it averaged 10.28 queries versus 16.00 for a method that removes individual records, with minimum-cost recovery of 50.6% versus 43.1%. JointCore tests groups of records before individual entries, addressing evidence that is redundant or useful only in combination. The researchers also introduced JointCoreBench, which supplies exact sufficient-set answers across nine interaction families and six workflow domains. Behavior preservation depended on the deterministic primary interface; alternative interfaces produced different behavior. The study does not establish deployment-ready memory compression.