Large-scale single-cell atlases have become indispensable resources for cell-type annotation and biological discovery. However, most existing annotation frameworks rely on static reference data and require re-accessing or retraining on previous datasets as new data emerge, which poses challenges for scalability, data sharing, and knowledge continuity. These methods are further constrained by catastrophic forgetting and batch-specific biases, limiting their ability to integrate knowledge across platforms, tissues, and modalities. Here we introduce scEvolver, a continual learning framework for single-cell annotation. scEvolver refines cell-type representations as class prototypes through memory-guided continual learning, incrementally accumulating knowledge without revisiting historical data. These online prototypes preserve intrinsic and consistent cell-type semantics across datasets while capturing informative within-class heterogeneity. Systematic evaluations demonstrate that scEvolver outperforms other methods in annotation accuracy, while requiring substantially fewer labeled reference samples for external query mapping. The framework maintains strong stability and generalization across diverse real-world scenarios spanning multiple platforms, tissues, and modalities. The application of scEvolver to inflammatory gut disease data reveals metaplastic transitions of epithelial cells, highlighting its potential to uncover context-specific cellular dynamics in complex disease settings.