Negative samples selection for contrastive learning is considerable in the field of sentence representation, especially for semantic textual similarity. Traditional in-batch negative sampling methods not only lack hard negative samples but also ignore potential false negative samples. Despite numerous methods trying to improve traditional sampling strategies, the challenge of consistently generating high-quality negative samples remains untackled. To address this pivotal issue, we propose the Cluster-Mined Negative Samples for Enhanced Unsupervised Sentence Representation Learning (CMNS) framework. Specifically, dynamic queues are utilized to store the K-means cluster samples, enabling the most appropriate selection of clusters to serve as negative samples. Additionally, we generate noise-based negative samples via stored clusters while simultaneously constraining potential false negative samples. Above all, CMNS employs clustering techniques to efficiently mine sufficient quantity of high-quality negative samples from unlabeled datasets. Extensive experiments illustrate that our approach not only overcomes the inherent limitations of traditional sampling methods but also improves the performance of sentence representations in downstream tasks, demonstrating measurable advancements over current methodologies. (codes and models are available at https://github.com/hamrain/CMNS ).