Modern high-field NMR spectrometers can stabilize the nominal sample temperature at a precision of less than 0.1 K. However, the actual sample temperature may differ from the nominal value by several degrees because the sample heating caused by high-power radio frequency pulses is not readily detected by the temperature sensors. Without correction, transfer of chemical shifts between different experiments causes problems in the data analysis. In principle, the temperature differences can be corrected by manual procedures but this is cumbersome and not fully reliable. Here, we introduce the concept of a "T-lock", which automatically maintains the sample at the same reference temperature over the course of different NMR experiments. The T-lock works by continuously measuring the resonance frequency of a suitable spin and simultaneously adjusting the temperature control, thus locking the sample temperature at the reference value. For three different nuclei, 13C, 17O and 31P in the compounds alanine, water, and phosphate, respectively, the T-lock accuracy was found to be <0.1 K. The use of dummy scan periods with variable lengths allows a reliable establishment of the thermal equilibrium before the acquisition of an experiment starts.