Agent repair memory reduces repeated Text-to-SQL failures

MERIT, a training-free agent described in a new arXiv paper, tries to stop language-model agents from relearning the same repair lessons in every episode. It stores both oracle-verified corrections and unsuccessful directions from completed Text-to-SQL repair attempts.

When a later task fails, the system classifies the failure type and retrieves relevant memories before the frozen model writes a revision. The method uses Qwen2.5-7B-Instruct with the same initial predictions and repair budgets, so the comparison focuses on memory use rather than model retraining.

The setup depends on benchmark feedback and is not a universal repair system. Still, it points to a practical design pattern for agents: durable memory should include what not to try, not only examples of successful fixes.