The messenger molecule nitric oxide (NO) is a key mediator of memory formation that can diffuse in the brain over tens of micrometres. It would seem therefore that NO derived from many individual neurones may merge into a volume signal that is inevitably ambiguous, relatively unspecific and thus unreliable. Here we report on the neuronal architecture that supports the NO−cyclic GMP signalling pathway in the mushroom body of an insect brain, the key centre for associative learning. We show that, in the locust (Schistocerca gregaria), parallel axons of intrinsic neurones (Kenyon cells) form tubular NO-producing zones surrounding central cores of NO-receptive Kenyon cell axons, which do not produce NO. This segregated architecture requires NO to spread at physiological concentrations up to 60 µm from the tube walls into the central NO-receptive cores. By modelling NO diffusion we show that a segregated architecture, which requires NO to act at a distance, affords significant advantages over a system where the same sources and targets intermingle. Segregation enhances the precision of NO volume signals by reducing noise and ambiguity, achieving a reliable integration of the activity of thousands of NO-source neurones. In a neural structure that forms NO-dependent associations, these properties of the segregated architecture may reduce the likelihood of forming spurious memories.