We are interested in learning models of non-stationary environments, which\ncan be framed as a multi-task learning problem. Model-free reinforcement\nlearning algorithms can achieve good asymptotic performance in multi-task\nlearning at a cost of extensive sampling, due to their approach, which requires\nlearning from scratch. While model-based approaches are among the most data\nefficient learning algorithms, they still struggle with complex tasks and model\nuncertainties. Meta-reinforcement learning addresses the efficiency and\ngeneralization challenges on multi task learning by quickly leveraging the\nmeta-prior policy for a new task. In this paper, we propose a\nmeta-reinforcement learning approach to learn the dynamic model of a\nnon-stationary environment to be used for meta-policy optimization later. Due\nto the sample efficiency of model-based learning methods, we are able to\nsimultaneously train both the meta-model of the non-stationary environment and\nthe meta-policy until dynamic model convergence. Then, the meta-learned dynamic\nmodel of the environment will generate simulated data for meta-policy\noptimization. Our experiment demonstrates that our proposed method can\nmeta-learn the policy in a non-stationary environment with the data efficiency\nof model-based learning approaches while achieving the high asymptotic\nperformance of model-free meta-reinforcement learning.\n