State Space Models (SSMs) have emerged as competitive alternatives to transformers in sequence modeling, particularly for processing long sequences. However, their ability to capture long-range dependencies faces limitations. We identify two main issues: local bias within a single SSM layer and over-smoothing when stacking layers. To address these, we propose Adaptive Jumping Knowledge (AJK), which combines representations from different layers. Integrating AJK into the Mamba architecture, we show improved performance in various training scenarios. AJK-enhanced models consistently outperform baseline models on public benchmarks.