Document-level relation extraction serves as a cornerstone for constructing structured knowledge. However, existing methods heavily rely on dense supervision signals and suffer from severe long-tail effects in distantly supervised scenarios characterized by sparse annotations and significant noise, resulting in restricted generalization capability on low-frequency relations and substantial performance degradation. To address these challenges, we propose PaMeRE, a partitioned memory method that enhances long-tail relation extraction through a dual memory mechanism: a general-purpose memory slot captures high-frequency relational patterns, while a dedicated long-tail memory slot focuses on modeling rare relations. The framework employs the Adaptive Fusion Gate(AFG) to achieve context-aware dynamic feature fusion, effectively integrating global semantic patterns with fine-grained relational features. Furthermore, we design a Dynamic Confidence Discrimination Loss that suppresses noisy signals through adaptive focal modulation and margin ranking constraints, prioritizing the identification of genuine relational patterns. Experiments on the ReDocRED benchmark demonstrate that PaMeRE achieves state-of-the-art performance under distant supervision (overall F1 score improved by 3% and long-tail relations F1 score improved by 4%).