Federated learning (FL) emerges as a promising approach for healthcare applications by combining machine learning with privacy-preserving techniques. However, two key challenges hinder its real-world applicability: client drift and client personalization. The former arises from non-IID data across clients, while the latter aims to tailor models to individual client data for improved local performance. Moreover, these challenges are inherently conflicting, i.e., optimizing for one can exacerbate the other. This paper proposes an innovative approach that integrates a hypernetwork with contrastive representation learning, which aims to shape the personality of each client under data heterogeneity. We evaluate the efficacy of our work through an experiment focused on Alzheimer’s disease diagnosis.