The challenge of reconstructing air temperature for environmental applications is to accurately estimate past exposures even where monitoring is sparse. We present XGBoost-IDW Synthesis for air temperature (XIS-Temperature), a high-resolution machine-learning model for daily minimum, mean, and maximum air temperature, covering the contiguous US from 2003 through 2023. XIS uses remote sensing (land surface temperature and vegetation) along with a parsimonious set of additional predictors to make predictions at arbitrary points, allowing the estimation of address-level exposures. We built XIS with a computationally tractable workflow for extensibility to future years, and we used weighted evaluation to fairly assess performance in sparsely monitored regions. The weighted root mean square error (RMSE) of predictions in site-level cross-validation for 2023 was 1.78 K for the minimum daily temperature, 1.19 K for the mean, and 1.48 K for the maximum. We obtained higher RMSEs in earlier years with fewer ground monitors. Comparing to three leading gridded temperature models in 2021 at thousands of private weather stations not used in model training, XIS had at most 60% of the mean square error for the minimum temperature and 93% for the maximum. In a national application, we report a stronger relationship between minimum temperature in a heatwave and social vulnerability with XIS than with the other models. Thus, XIS-Temperature has potential for reconstructing important environmental exposures, and its predictions have applications in environmental justice and human health.