Active Inference with Reusable State-Dependent Value Profiles