Offline Reinforcement Learning with Value-based Episodic Memory