Disentangling Memory and Reasoning Ability in Large Language Models