Memory-efficient Reinforcement Learning with Value-based Knowledge Consolidation