Lifelong Person Re-Identification (LReID) is formulated as a continual learning task, where models must retain discriminative power on previously seen domains while adapting to new domains with emerging data. Current LReID studies focus excessively on preserving old knowledge, lacking effective strategies to guide the acquisition and integration of new knowledge from current tasks, which causes a plasticity-stability imbalance and impairs overall performance. To address this challenge, we propose a Bidirectional Knowledge Distillation Network (BKDNet) for LReID. Specifically, an old-domain stability knowledge constraint module is designed in BKDNet, which constructs a robust knowledge repository for the old domain, and combines the advantages of both an online encoder and a momentum encoder to achieve stable distillation of old knowledge, effectively preserving historical information. Meanwhile, to balance stability and plasticity, a current-domain guidance module is proposed to build an expert model that facilitates the main model to adapt to new data and distribution changes. Furthermore, an old-domain sample augmentation strategy is designed, which trains a predictive network to generate synthetic samples that augment replay data, thereby enhancing the model’s resistance to forgetting and reducing storage overhead. Experimental results demonstrate that the overall performance of our method significantly exceeds state-of-the-art unsupervised LReID methods.