Our goal is to develop silicon learning systems. One impediment to achieving this goal has been the lack of a simple circuit element combining nonvolatile analog memory storage with locally computed memory updates. Existing circuits [63, 132] typically are large and complex; the nonvolatile floating-gate devices, such as EEP ROM transistors. typically are optimized for binary-valued storage [17], and do not compute their own memory updates. Although floatinggate transistors can provide nonvolatile analog storage [1, 15], because writing the memory entails the difficult process of moving electrons through Si02 , these devices have not seen wide use as memory elements in silicon learning systems.We have fabricated synapse transistors that not only possess nonvolatile analog storage, and compute locally their own memory updates, but also permit simultaneous memory reading and writing, and compute locally the product of their stored memory value and the applied input. To ensure nonvolatile storage, we employ standard floating-gate MOS technology, but we adapt the physical processes that write the memory to perform a local learning function. Although the Si02 electron transport still is difficult, and does require high voltages, because our devices integrate both memory storage and local computation within a single device, we expect them to find wide application iil silicon learning systems.We call our devices synapse transistors because, like neural synapses [11], they compute the product of their stored analog memory and the applied input. Also like neural synapses, they can learn from the input signal, without