Self-adaptive gating in associative memory networks

Gating—adaptive, dynamical control of effective neuronal timescales—underpins the success of gated RNNs (LSTM/GRU) and has clear biological analogs (e.g., shunting inhibition, neuromodulatory control). While its role in random RNNs has recently been clarified, its impact on associative memories remains unclear. We study a continuous-state Amari–Hopfield-type network augmented with a self-adaptive gate that multiplicatively modulates each unit’s timescale based on the current network state. The gated model exhibits a retrieval phase far beyond the classical storage limit, effectively bypassing the order-to-spin-glass transition present in ungated networks. This capacity gain does not arise from narrowed basins; rather, it trades off with reduced fidelity (smaller steady-state overlaps). Moreover, gating generically produces multistable “clouds” of attractors corresponding to continuously varying overlaps set by initial neuronal activities. We corroborate direct many-body simulations with a dynamical mean-field theory. Our results provide a minimal nonequilibrium mechanism that enhances capacity and reshapes attractor geometry, offering new conceptual avenues for associative memories in both biological and artificial network circuits.