Skillful hydroclimate forecasts on longer timescales are crucial for decision-makers and resource managers to mitigate climate-driven natural disasters, such as multi-year drought, agricultural losses, and an increase in wildfire dangers. Decadal climate prediction based on the Earth system model with proper initialization can provide skillful hydroclimate predictions in the western US. However, the most decadal climate predictions still suffer from large model biases and initialization shocks, which can severely contaminate model forecasts and may substantially reduce predictive skills, particularly for non-Gaussian distributed variables, such as rainfall, streamflow, groundwater, and wildfire. In this presentation, I will show that our drift-free decadal climate prediction system has demonstrated the multi-year predictive skill of soil water anomalies in the western US. This system consists of an ocean data assimilation approach using the coarse resolution of a fully coupled Community Earth system model (CESM1.0) and significantly reduces the model climate drift during prediction by adjusting the model biases in the mean state and the climate response to the radiative forcing during the assimilation process. The detailed approach in our drift-free prediction system and its application on the Earth system predictability will be discussed.