Graph contrastive learning (GCL) has garnered increasing attention due to the ability to handle unlabeled data, which are particularly prevalent in the real-world graph-structured data. The contrastive mode of GCL necessitates that the graph encoder gives node embeddings preserving consistent features. Most existing graph encoders adopt message passing mechanism and operates under the homophily assumption where connected nodes predominantly belong to the same class. However, this mechanism suffers from local structure over-reliance, leading to noise accumulation during information aggregation on heterophilic graphs. Therefore, capturing global relationships while suppressing the data contamination remains a critical challenge. Although recent studies have attempted to employ a sequence model Mamba for graph encoding to address this challenge, existing graph node sequencing methodologies demonstrate insufficient alignment with Mamba’s position-aware processing paradigm. In this paper, we propose HambaGCL, a heterophilic GCL model based on our graph encoder called Multi-Hop Mamba (Hamba). This newly proposed Hamba Encoder translates the multi-hop structural information into Mamba’s expected position-aware input sequences. Meanwhile, Hamba employs tree decomposition to build non-overlapping hop sequences, thus avoiding information contamination. Moreover, experiments across five heterophilic benchmark datasets reveal that HambaGCL achieves significant improvements over state-of-the-art models.