Plateau-gated one-shot plasticity supports continual recognition memory

Biological memory systems store single experiences while continuing to learn, but how one-shot plasticity limits interference with existing memories is unclear. Behavioral timescale synaptic plasticity (BTSP) rapidly modifies synapses active within seconds of a dendritic plateau. We isolate its plateau-triggered component in a model where plastic weights and a stable instructive pathway jointly determine whether a plateau occurs, closing a feedback loop between the synaptic state and the plastic event that modifies it. For unstructured inputs, instantiated by independent uniform signed patterns, the dynamics reduce exactly to pathway alignment, which determines the fidelity of the instructed representation, the synaptic-turnover rate and the mean rewrite interval. For structured inputs, instantiated by correlated bimodal Curie–Weiss patterns, the instructive pathway biases which component of input structure enters the plastic synaptic state. In a BTSP-inspired continual-recognition network, combined instructive and plastic drives determined the selected memory unit for each one-shot write, whereas a Hebbian control used the same plastic weights for credit assignment and memory storage. The BTSP-inspired network remained accurate at longer repeat lags than Hebbian controls, an advantage that grew with network size, with both architectures optimized independently at every repeat lag. A reduced theory predicted held-out accuracy, lag capacity and dynamics of memory-trace strength directly from optimized parameters. It showed why intermediate proximal and distal coupling was optimal: proximal plastic drive guided plateau generation toward selected memory units, slowing synaptic turnover but limiting new encoding, whereas distal instructive drive enhanced familiar responses but could also make novel inputs appear familiar. The memory-trace strength in the rate-and-depth-matched Hebbian control still decayed faster and showed less effective credit assignment than in the BTSP-inspired network. These results connect dendritic plateau physiology to continual memory and support partial separation of allocation from storage as a mechanism for limiting interference during continual learning.