Objective: Existing methods for continuous relation extraction (CRE) face remarkable challenges in complex Chinese semantic environments, particularly the problem of catastrophic forgetting when learning new tasks while retaining knowledge from old tasks. Traditional approaches often retrain models on a combination of historical and new data, leading to inefficiencies and resource constraints as data volumes increase. To address these limitations, this study aims to develop a robust CRE model that can efficiently learn new knowledge while preserving historical relationships, even in scenarios with complex sentence structures, overlapping entities, and imbalanced data distributions. The proposed model integrates prototype representations, memory replay strategies, and contrastive learning to enhance feature discrimination and stability in embedding spaces, thereby improving the classification performance across single-domain and cross-domain datasets. Methods: The proposed prototype-based continuous complex relation extraction network (PBCRE-Net) model consists of two primary stages: initial training and memory replay, designed to mitigate catastrophic forgetting and improve adaptability in dynamic learning environments. The initial training includes: 1) Entity-aware feature extraction: Input texts are processed using a pretrained BERT model to generate contextual embeddings with special tokens ([E11], [E12], [E21], and [E22]) that mark entity boundaries. 2) Supervised contrastive learning: A dual-head classifier (classification and contrast) is employed to minimize intraclass distances and maximize interclass distances in the embedding space. This objective is achieved through a combination of cross-entropy loss and contrastive loss. 3) Prototype generation: For each relation category, representative samples are selected via K-means clustering and their prototypes are computed as weighted averages of cluster centroids to capture category-specific features. The memory replay includes: 1) Memory sample selection: Memory modules store exemplars from previous tasks using K-means clustering. Weights are assigned based on cluster distribution to balance sampling during replay. 2) Memory augmentation: To prevent overfitting, synthetic samples are generated by swapping entity pairs or appending unrelated sentences to existing exemplars, thereby expanding the memory pool. 3) Consistency Loss: During replay, an embedding consistency constraint is applied to maintain stability in the embedding space across tasks. 4) Joint optimization: The model is trained on a mixture of new task data, historical memory samples, and augmented samples, combining cross-entropy loss and consistency loss. Results: Experimental evaluations on CMeIE (medical domain) and ASaRED (alloy domain) datasets demonstrate the superiority of PBCRE-Net in complex CRE scenarios. Single-domain Performance: On the CMeIE dataset, PBCRE-Net achieved an average accuracy (ACC) of 84.27% and a macro F1-score (Macro-F1) of 81.93% across 10 incremental tasks. Notably, the proposed model outperformed baseline models, such as EMAR and CRL, by 3%-5% in subsequent tasks (T8—T10), where catastrophic forgetting is highly severe. The model effectively handled triplet overlap (e.g., entity-pair and single-entity overlaps) and class imbalance, an objective accomplished through prototype-based contrastive learning and memory augmentation. Cross-domain adaptability: In cross-domain experiments combining CMeIE and ASaRED, PBCRE-Net maintained an ACC of 74.16% and Macro F1 of 69.34% across 10 tasks, considerably surpassing competing models (e.g., CRECL and DPCRE). The memory replay mechanism and consistency loss ensured stable embedding spaces despite domain shifts, thus reducing catastrophic forgetting in critical relation categories like material composition and alloy property. Robustness to memory constraints: Reducing the memory size from 20 to 5 samples per task decreased the performance of the proposed model by 15%, yet PBCRE-Net outperformed the alternatives under constrained memory conditions. This highlights its efficiency in real-world scenarios with limited storage. Conclusions: This study introduces PBCRE-Net, a novel CRE framework that mitigates catastrophic forgetting through prototype representations, memory replay, and contrastive learning. Key contributions include a supervised contrastive learning strategy to enhance feature discriminability in complex semantics and a memory augmentation mechanism to mitigate overfitting and stabilize embedding spaces. Superior performance in single-domain and cross-domain CRE tasks was validated by extensive experiments. Future work will extend PBCRE-Net to multilingual settings via cross-lingual transfer learning and incorporate physical constraints to improve relation extraction accuracy in scientific domains. In addition, addressing polysemy through semantic alignment techniques will further enhance the applicability of PBCRE-Net.