One of the most fundamental tenets in neuroscience is that information is stored in the strength of synapses, and that learning relies on changes in synaptic strength. Here we aim to extend this tenet into the temporal domain by proposing that synapses not only learn to strengthen or weaken, but can learn when to be weak or strong on the subsecond scale. It has been known for more than half a century that synapses exhibit short-term synaptic plasticity (STP): a form of presynaptic, use-dependent, plasticity in which consecutive spikes can elicit progressively depressing (short-term depression/paired-pulse depression) or facilitating (short-term facilitation/paired-pulse facilitation) postsynaptic potentials. Here we provide the first experimental evidence that STP may be learned. By pairing postsynaptic activity to co-occur with the first or last spikes of a presynaptic train, we shifted STP towards depression or facilitation, respectively. This result is consistent with our prediction that short-term synaptic dynamics are tuned to peak at the time of postsynaptic activity. Finally, we show that when Learned-STP is incorporated into artificial neural network models, the STP parameters can be trained to generate interval selective neurons, solve the spatiotemporal XOR task, as well as a complex temporal Morse code recognition task in the absence of any recurrency. Together these studies establish a novel experimental and theoretical framework to potentially extend our understanding of synaptic plasticity and synaptic computations into the temporal domain.