Rivers in the Arctic are drawing significant attention as the region has warmed four times faster than the global average. River temperature affects almost all physico-chemical equilibria, which in turn control biogeochemical cycles, biological behaviors, and many aspects of water quality. River temperature holds significant socio-economic implications for fisheries and food production for local communities. However, little work has been done to forecast river temperature dynamics in the Arctic to understand the impacts of warming on riverine ecosystems, primarily due to limited data and non-standardized methods. The newly-released AKTEMP dataset and recent successful applications of machine learning offers an opportunity to develop scalable, transferable river temperature models that can aid in these projections. Here, we developed a robust, physics-informed Long Short-Term Memory (LSTM) model based on the recently curated AKTEMP stream temperature database to better predict river temperatures and elucidate the key processes driving river temperature dynamics in ungauged snow- and permafrost-affected basins in Alaska. Our reported evaluation metrics over the study domain (NSE=0.87, RMSE=0.99 °C, Pearson-R=0.97) outperformed the best available models of river temperature in the literature for high-latitude regions. Our findings support the idea that river temperature is an integrated function of various hydrometeorological forcings across a broad spectrum of timescales. These range from days for air and soil temperature to months for soil moisture and snow, varying across different hydrological conditions. We also found that glaciers and permafrost play vital roles in seasonal river temperature predictions. Our model enables future applications to unravel river temperature regime shifts under a changing climate.