Accurate site-level identification of RNA post-transcriptional modifications is essential for deciphering epitranscriptomic regulation. However, existing deep learning predictors typically rely on either convolutional filters with limited receptive fields or self-attention mechanisms with quadratic complexity, while parallel hybrid networks often struggle to coordinate global sequence context with local motif sensitivity. Here, we present HMB-Net, an interpretable hybrid framework that couples a Selective State Space Model (Mamba) with a deep Bidirectional Long Short-Term Memory (Bi-LSTM) network in a sequential “Latent-to-Motif” hierarchy. Rather than extracting features in parallel, HMB-Net first employs a Mamba encoder to capture long-range sequence dependencies in linear time, which then guides a multi-layer Bi-LSTM to resolve fine-grained motif boundaries, followed by a bidirectional cross-attention module to integrate multi-scale representations. Comprehensive benchmarking across 20 mammalian datasets spanning $N^6$-methyladenosine (m${}^6$A), 5-methylcytosine (m${}^5$C), and $N^1$-methyladenosine (m${}^1$A) modifications in human, mouse, and rat demonstrates that HMB-Net consistently outperforms representative convolutional, recurrent, and Transformer-based baselines. Ablation experiments further confirm the benefit of the global-to-local encoding order over single-branch, reverse-ordered, and parallel dual-branch variants. Moreover, dimensionality reduction and integrated-gradient saliency analyses show that HMB-Net learns well-separated feature representations and autonomously focuses on conserved flanking consensus motifs (such as RRACH) without predefined biochemical rules, providing a reliable computational tool for transcriptome-wide modification mapping.