Lossless hyperspectral image (HSI) compression poses significant challenges due to complex spatial–spectral correlations and the massive data volume. We propose a learning-based compression framework that introduces an Intra-Slice Adaptive Clustering (ISAC) module to partition each sub-image based on pixel intensity distributions. This clustering reduces intra-region variability and enhances the accuracy of probability modeling. To further exploit structural dependencies, we design a Multi-Layered Context Modeling Network (MLCM-Net), which integrates three modules: (1) a forward inter-slice context module (FISCM) that conditions current predictions on the previous spectral slice, (2) a bit-plane context module (BPCM) that leverages high-bit information to guide low-bit compression, and (3) a spatial location context module (SLCM) that captures local spatial similarity for more effective modeling. Experimental results on five benchmark HSI datasets demonstrate that MLCM-Net achieves superior lossless compression performance compared to both traditional and learning-based baselines, while maintaining low complexity. The proposed method also demonstrates strong robustness across common HSI scenarios with diverse spatial sizes, spectral bands, and scene semantics. The code will be released at https://github.com/LSJ047/ISAC-HIC .