Adapts the agent learning speed per domain based on prediction accuracy, surprise, and errors. Fast learning when predictions fail, slow consolidation when accurate — the agent learns how fast to learn.