The vestibulo-ocular reflex (VOR) is under long-term adaptive regulation to minimize retinal image slip during head movement; normally this process keeps VOR gain (eye velocity divided by head velocity) near 1.0. It has been common to think of the adaptive mechanism as a single pure gain element, although some properties of the system (e.g., frequency-selective changes in the gain of the VOR) argue that it must be more complex. We now report new observations on the frequency selectivity of the adaptive mechanism. Our data suggest a new model in which the VOR operates as a series of parallel, temporal frequency channels, each of which has an independently adjustable gain element. Adaptive changes were produced by oscillating monkeys sinusoidally at a single temporal frequency (0.2 or 2.0 Hz) in visual conditions that cause either increases (toward two) or decreases (toward zero) in VOR gain. When tested in darkness at the adapting frequency, the VOR showed large changes in gain and little or no change in phase. When tested at frequencies other than the adapting frequency, the VOR showed less pronounced changes in gain and unexpected changes in phase. The phase changes were orderly but depended in a complex way on adapting frequency, testing frequency, and VOR gain. We have tested the channels concept by calculating the response properties of a mathematical model that processed its inputs in parallel pathways. The model reproduced our data when we assumed that the vestibular primary afferents were distributed in an orderly way to parallel brain channels that had differing dynamics: vestibular inputs with more phase lead projected to higher frequency channels, which themselves had faster dynamics than their low frequency counterparts. Such an organization, when regulated by an adaptive controller that can selectively alter the gain of one channel, could play a key role in establishing and maintaining the frequency-independent performance seen in the adult VOR.