Substantial research has been devoted to oracle character recognition. However, the inherent long-tailed distribution in these datasets remains a persistent challenge, undermining the performance. To address the issue, this work proposes a Dual Manifold Volume-Balanced Framework (DMVBF) for the long-tailed oracle character recognition. We first show conventional models exhibit significant manifold volume skewness: head classes occupy far larger feature space than tails, biasing decision boundaries toward dominant categories. Building on this insight, we integrate manifold volume with Logit Adjustment (LA) and Block-wise Knowledge Distillation (BWKD) to develop dual balanced techniques, called MVB-LA and MVB-BWKD, which are then embedded within a Probabilistic Contrastive Learning (ProCo) method. Our approach employs a two-stage training framework: pre-training with ProCo for half the epochs, followed by iterative refinement via the dual balanced techniques. By leveraging manifold volume as a weighting factor in MVB-LA and a block segmentation criterion in MVB-BWKD, manifold volumes are promoted to balance across classes, thereby equalizing their representation spaces and alleviating long-tailed bias. Extensive experiments on three long-tailed datasets—including two oracle character datasets—demonstrate that our method achieves superior performance compared to state-of-the-art baselines.