Document-level relation extraction in document understanding aims to identify complex entity relationships across sentences, but existing methods often face challenges when capturing long-distance entity dependencies. To address this issue, this paper proposes a document-level relation extraction model based on Entity Dependency Memory Enhancement (EMDRE). By introducing memory-enhanced technology from the computer vision domain, token Turing machine (TTM), the model effectively improves the ability to model long-distance entity dependencies in document understanding tasks. This approach enhances the memory mechanism, significantly improving the model’s ability to capture complex entity relationships within documents. Additionally, to tackle the issue of class imbalance in document-level relation extraction tasks, this paper presents an adaptive ranking optimization strategy that dynamically adjusts the contribution of positive and negative samples to model training, further boosting model performance. Experimental results demonstrate that the EMDRE model outperforms existing baseline models in document-level relation extraction tasks, validating its effectiveness and innovation in document understanding.