MINTEval benchmark for AI memory systems accepted to NeurIPS 2026

The MINTEval team says its benchmark was accepted to NeurIPS 2026. It tests how AI agents and memory systems handle changing information, including Wikipedia revisions and GitHub commits. The team reports that seven representative systems averaged 27.9% accuracy, with the best reaching 33.4%. It attributes the difficulties mainly to problems with retrieval and memory construction.