Personalized federated recommendation (PFR) enhances the recommendation performance of client models by personalizing the modeling of heterogeneous data. Existing methods achieve personalization by incorporating local item embeddings on global item embedding. However, two core challenges persist. First, they lack fine-grained differentiation in data heterogeneity across datasets, leading to inconsistent performance when applying the same modeling approach. Second, solely modeling global knowledge through shared item embeddings is insufficient, reducing model performance, as client selection further exacerbates knowledge forgetting. In this work, we highlight the dataset heterogeneity and observe the impact of this heterogeneity and knowledge forgetting on model performance. We propose an Adaptive Personalized Federated Recommendation through Global Knowledge Distillation (APFR-GKD) to address these challenges. This approach introduces an adaptive personalization mechanism that adjusts the algorithm’s personalization level based on the degree of data heterogeneity. Additionally, we enhance global knowledge integration by sharing item embeddings and scoring functions, incorporating global historical models, and guiding local model updates through knowledge distillation. Comprehensive experiments on four datasets demonstrate that our method consistently outperforms baseline algorithms on heterogeneous datasets. Ablation studies further validate the effectiveness of our approach.