In dynamic environment, Learning Classifier System (LCS) evolves classifiers to fit the current situation, but may forget classifiers which were useful for previous situations. Our main idea is that, we store the forgotten classifiers as archives and generate new classifiers by recombining them to fit the current situation. Specifically, we propose an archive-based LCS called Arc-XCS, which detects environmental changes and generates classifiers based on the archive. The experimental results on the benchmark problem show that, Arc-XCS successfully stored good classifiers when each environmental changes occurs; compared to the conventional LCS (XCS), Arc-XCS reaches better performances with fewer trainings.