Learning to Retrieve: Dual-Level Long-Term Memory for Text-to-SQL Agents

In three linesMERIT, a multi-horizon memory retrieval framework for text-to-SQL agents, maintains episode-level and turn-level memory with learned retrieval policies optimized via reinforcement learning. On BIRD-Interact, it outperforms no-memory and static-retrieval baselines, improving success rate while reducing interaction turns.

Summary generated by Claude — human-verified