DISCERN preprint reports gains in resolving conflicting AI agent memories

Researchers report that DISCERN, an AI agent memory framework, improved answer correctness and decisions to decline unsupported answers in benchmark tests. The preprint reports abstention F1 of 0.924 on the synthetic ENT-310 enterprise sales benchmark, versus at most 0.802 across eight baselines. DISCERN preserves conflicting and superseded facts in a graph, then resolves entities and time states when answering questions. Macro answer correctness reached 0.861 versus 0.806 for the strongest baseline. Without tuning, the pipeline also led macro answer correctness on the full LoCoMo corpus, scoring 0.663 versus 0.638. Evaluation used a council of four language-model judges.