<p>Non-exemplar class-incremental learning has attracted increasing attention in recent years due to challenges related to data privacy and memory constraints. However, the inability to retain previous samples in this setting presents a significant obstacle to preserving past knowledge and mitigating catastrophic forgetting. Existing methods often attempt to address feature space shifts through prototype augmentation. Yet, relying on a single prototype to represent an old class fails to capture the full complexity of its feature distribution, reducing the model’s robustness to distributional changes. Furthermore, continually fine-tuning the classifier across tasks tends to overwrite previously learned knowledge, resulting in severe classification bias–especially when prototype representations are limited. To address these challenges, we propose a feature adaptation method for non-exemplar class-incremental learning. It comprises two key components: dynamic prototype calibration and progressive classifier adaptation. Dynamic prototype calibration adaptively adjusts the feature representations of old classes to align with the evolving feature space during incremental learning, effectively enhancing the model’s ability to retain prior knowledge. Meanwhile, progressive classifier adaptation decouples the learning of old and new class classifiers to prevent mutual interference and reinforces old class learning through a transfer-then-merge strategy, preserving original knowledge while continually improving the classifier’s adaptability to the dynamically evolving feature space. Finally, extensive experiments demonstrate the effectiveness of our method, showcasing that it achieves state-of-the-art performance.</p>