Abstract: A model based on dynamic memory and progressively frozen classifiers (DMPFC) was proposed to address the catastrophic forgetting issue in continual relation extraction, particularly the challenges of insufficient memory allocation and classifier parameter drift. By dynamically adjusting memory allocation and gradually freezing the classifier parameters corresponding to each task, the model's ability to retain historical knowledge over the long term was effectively enhanced. In terms of methodology design, the model mainly included two stages: new relation learning and memory replay. During the learning stage, a data augmentation strategy was adopted along with a progressively frozen classifier mechanism, where an independent classifier was allocated for each new task to achieve representation isolation. Meanwhile, memory space was dynamically allocated based on the semantic importance of relations, representative samples were selected through hierarchical clustering, and relation prototypes were updated using a moving average strategy. In the replay stage, dual-view contrastive learning and hierarchical knowledge distillation were combined to enhance representational consistency and knowledge retention capabilities. Additionally, a hybrid prototype similarity metric was introduced to improve the robustness of the inference process, thereby jointly optimizing feature representation and knowledge preservation. Experimental results demonstrate that the DMPFC achieves accuracies of 87.5% and 81.3% on the final tasks of the FewRel and TACRED datasets, respectively, significantly surpassing existing mainstream methods, which proves its effectiveness in the continuous recognition and updating of relations within complex contexts. This study provides a feasible technical pathway and an algorithmic foundation for constructing practical continual learning systems in scenarios involving dynamically evolving and large-scale knowledge graphs.