The self-supervised point cloud completion models are capable of training without the need for paired or complete point cloud data, relying solely on the incomplete point clouds themselves. However, it is precisely this advantage that limits the available information of the models, hindering their performance. Additionally, existing self-supervised methods focus more on designing consistency between partial point clouds and their various projection variants, while neglecting to explore complementary information and prior knowledge within the partial point clouds themselves. In order to address the challenges mentioned above while maintaining the advantages of self-supervised models, we propose a method that provides 2 kinds of additional prior knowledge to the model. The first is a Mutual Learning Memory Bank (MLMB), which enables mutual learning among different partial point clouds in the dataset to guide point cloud completion. The second is the 3D Planar Symmetry Detection (PSD) module, which is capable of extrapolating missing information about an object by utilizing its inherent symmetrical properties. In addition, we incorporated a weighted Chamfer Distance loss to compare the newly introduced prior information with the input partial point cloud, which enhances the model's ability to distinguish between the differences and consistencies in content between the prior information and point cloud. Experiments demonstrate that our method achieves state-of-the-art among self-supervised models and unsupervised methods.