Towards Teachable Reasoning Systems: Using a Dynamic Memory of User Feedback for Continual System Improvement