<p>In online teaching scenarios, automatic assessment of student engagement suffers from limited real-time capability, high computational cost, and redundant information in long-duration videos. To address these challenges, a student engagement recognition model driven by prediction error–based memory updating (DMDN) is proposed. A multi-stage framework with feature decoupling is adopted. A behavior dynamics branch is used to extract coarse-grained behavioral features. A facial attention branch is used to extract fine-grained facial representations. Semantic interference during heterogeneous visual feature modeling is thus avoided. Furthermore, a dynamic memory updating module based on prediction error is introduced. The deviation between current observations and historical memory is characterized. The memory update strength is adaptively adjusted. Selective encoding of key behavioral segments is achieved. Interference from redundant temporal information is effectively suppressed. The model parameters and deployment scheme are revised as follows. A two-stage deployment strategy is adopted, in which backbone features are extracted and cached offline, decoupling heavy feature extraction from online inference. Only a lightweight inference backend with 3.18M trainable parameters is executed during real-time prediction. Experimental results on the DAiSEE dataset demonstrate that a Top-1 accuracy of 64.65% is achieved, validating the effectiveness of the proposed method under strict computational constraints. This design demonstrates its potential in resource-constrained scenarios. Ablation studies further verify the effectiveness of the feature decoupling structure and the dynamic memory update mechanism in improving performance.</p>