Memory-efficient Reinforcement Learning with Knowledge Consolidation