Assembly-based computations through contextual dendritic gating of plasticity.

Neuronal assemblies-groups of strongly connected neurons-are widely considered fundamental building blocks of perception and memory. While experimental evidence supports their existence, it remains unclear how assemblies can be learned, maintained, and combined to support complex computations without mutual interference. We present a biologically grounded framework in which nonlinear dendritic branches and functionally distinct inhibitory populations drive the formation and combination of overlapping assemblies across multiple areas. By gating synaptic plasticity at the level of individual dendrites, context-dependent disinhibition allows the same neurons to participate in multiple assemblies without overwriting stored representations. Crucially, this stability enables assemblies to be projected and associated across brain areas, providing a circuit-level substrate for compositional, multi-area computations. We demonstrate these principles in a visual-auditory association task, where context-specific assembly representations separate ambiguous inputs. Together, our results show how dendritic and inhibitory circuit mechanisms support robust assembly-based computation in biologically realistic networks.