With privacy regulations in artificial intelligence (AI) gaining increasing attention, machine unlearning (MU) has emerged as a critical technique in privacy-preserving AI, allowing individuals to remove the influence of their data from pre-trained models. Exact unlearning, the gold standard, involves discarding target data and retraining the model from scratch. Although theoretically optimal, exact unlearning is computationally intractable for large-scale models. Approximate unlearning methods optimize model weights using full or partial datasets, offering a practical trade-off but risking regulatory violations by reusing data marked for forgetting. To address this, we propose Neighbor-Collaborative Forgetting and Regret Updating (NCFRU), a novel MU framework that avoids retraining on forgotten data. NCFRU partitions the dataset into retained and forget sets, using retained samples to form a “forget-support set” with neighborhood relationships. The NCFRU operates through an iterative mechanism. Initially, during the collaborative forgetting phase, gradient ascent combined with feature-space disentanglement penalties is employed to weaken the statistical dependencies of forget-support samples within local manifolds. The regret updating step then reconstructs the local topological structures, establishing new connectivity patterns. Extensive experiments across diverse datasets demonstrate that NCFRU optimally balances computational efficiency and regulatory compliance.