This paper proposes a novel frequency-calibration approach based on an adaptive neural–fuzzy inference system (ANFIS). In normal mode, an oven-controlled crystal oscillator (OCXO) is steered by integrating a time-interval counter, a fuzzy controller, a digital-to-analog (D/A) converter, and other components such that its frequency can follow a primary cesium atomic clock. In addition, under the effects of aging on the OCXO and the variation in ambient temperature, the control messages are collected to train an ANFIS in this mode. When the system enters holdover mode, the control messages predicted by the ANFIS are used to steer the OCXO to maintain its performance within the tolerance required by user applications. Experimental results indicate that the frequency stability of the OCXO can be improved from a few parts in $10^{10}$ to $10^{13}$ for an average time of one day in normal mode, and its frequency stability can be maintained within a few parts in $10^{12}$ over a measurement period of one day in holdover mode.