Industrial anomaly detection based on computer vision has gained significant attention. The feature memory bank, as a key research direction in this field, has been extensively studied in recent years. Traditional memory bank-based methods are characterized by one-to-one learning strategies, neglect of positional information, and lack of utilization of prior knowledge about anomalies. As a result, these methods exhibit high system complexity and are ineffective in detecting logical anomalies and near-distribution anomalies. To address these challenges, we propose a Unified Memory Bank for Discerning Near-distribution Anomalies (UM-DNA). This method embeds explicit positional information to integrate spatial context into features, and constructs dual memory banks to enable a unified model. Additionally, UM-DNA introduces a near-distribution anomaly synthesis technique to leverage prior knowledge of anomalies and employs a discriminative focus loss function to train the reconstruction network. These innovations significantly enhance the detection performance for near-distribution anomalies. Compared with other methods, UM-DNA demonstrates exceptional effectiveness and robustness on the MVTec AD and VisA datasets.