Knowledge graph reasoning aims to find the missing links in knowledge graphs and is an important fundamental task. Existing methods mostly reason end-to-end and ignore the prior knowledge in the knowledge graph. In this paper, we attempt to mine prior knowledge from the knowledge graph based on counterfactuals and to use the prior knowledge to enhance the model. Specifically, we begin by constructing counterfactuals to assign a weight for each relation as prior knowledge and then perform reasoning based on both prior knowledge and reinforcement learning. This approach combines the advantages of prior knowledge and neural networks. Experiments on three large datasets show that the prior knowledge extracted from counterfactuals is effective in improving the multi-hop reasoning model. Prior knowledge also has the advantage of being path-length independent, which mitigates the performance degradation in multi-hop reasoning when the reasoning path is excessively long.