<p>Unsupervised person re-identification aims to learn highly generalizable pedestrian representations without relying on manual annotations. Existing memory bank-based contrastive learning methods primarily depend on clustering to generate pseudo-labels and train the model through dynamic memory updates. However, their performance is constrained by the inter-class merging phenomenon, where dominant clusters gradually absorb weaker ones during training, leading to the accumulation of pseudo-label noise. Conventional approaches often regard boundary features as noisy samples, overlooking their potential value. Through systematic experiments, we identify and define a unique type of boundary feature–Volatile Features–whose dynamic variability plays a crucial role in inter-class merging and serves as a major source of pseudo-label noise. To address these challenges, we propose the Multicluster-Guided Contrastive Learning of Volatile Features (MGVF) framework, centered on Volatile Features. The framework introduces the Multicluster Centroid Modulator (MCM), which uses a dual-parameter clustering algorithm to identify Volatile Features and dynamically adjust sample weights to cluster centroids, enhancing the use of boundary information. Additionally, we present the Dynamic Global Feature Update Strategy (DGF) to address the limitations of traditional update methods that struggle with inter-class variations and contextual relationships. Finally, we introduce the Volatile Feature Exploration (VFE) loss to further capture latent information in Volatile Features. By reinforcing the learning of Volatile Features, our framework effectively mitigates inter-class merging issues during training. Extensive experimental results demonstrate that MGVF achieves state-of-the-art performance on multiple benchmark datasets for unsupervised person re-identification, thereby validating the effectiveness of the proposed approach.</p>