Memory Architectures for Multi-Turn Text-to-SQL: A Benchmark and Empirical Study

This research introduces EnterpriseMem-Bench, a novel multi-turn Text-to-SQL benchmark with 300 sessions and 1,400 turns from enterprise domains. It empirically evaluates five frontier models, including GPT and Claude variants, revealing that stateless multi-turn Text-to-SQL models achieve zero execution accuracy by Turn 3.