<p>Current lifelong person re-identification (LReID) methods fail to adequately address the critical issue of catastrophic forgetting caused by parameter updates and cross-domain distribution shifts, particularly neglecting the important distinctions between inter-person and intra-person feature distributions. Methods based on data replay and knowledge distillation also encounter limitations in data privacy preservation and forgetting resistance performance. We propose a Distribution-Optimized Knowledge Update (DOKU) framework to tackle the challenges in lifelong person re-identification. It introduces an instance-level distribution module that models the feature distribution of each instance, thereby capturing local pedestrian variations across different domains. The resulting distribution-aware prototypes not only preserve identity-specific central information but also encode intra-class diversity, providing richer fine-grained knowledge for lifelong learning. We propose a joint distribution optimization module that aligns feature distributions across source and target domains via cross-domain joint distribution learning, mitigating inter-domain discrepancies through global distribution modeling. A distribution-oriented knowledge update module is developed to transform instance-level predictions into multivariate Gaussian distributions. This module dynamically balances knowledge retention and absorption, enhancing feature representations of new identities via prototype-guided distribution learning to achieve cross-domain adaptation, all without storing raw data. Extensive experiments on multiple benchmark LReID datasets demonstrate that the proposed method significantly improves both forgetting resistance and generalization performance compared to conventional state-of-the-art approaches. Specifically, the proposed approach achieves 5.2% higher mAP and 6.3% better Rank-1 accuracy on seen datasets, along with approximately 3% improvement in both metrics on unseen datasets.</p>