Balancing Stability and Flow in Hippocampal Networks via Inductive Bias and Learned Symmetry Breaking

The hippocampus must balance stable memory representations with internally generated sequential dynamics underlying replay and prediction. How hippocampal circuitry achieves both remains unclear. Here, we show that recurrent neural networks trained on prediction tasks converge to a mixed-symmetry dynamical regime, in which dominant symmetric recurrence stabilizes an attractor while a weaker antisymmetric component induces directed flow. This structure accelerates learning and supports robust replay and prediction. We further show that such symmetry breaking can arise from biologically plausible spike-timing-dependent plasticity (STDP) rules, yielding a tilted Mexican-hat connectivity profile without explicit architectural constraints. Initializing networks with CA3-like structured connectivity biases learning toward this regime, and improves optimization efficiency and performance. These results suggest that hippocampal computation reflects an interaction between biologically constrained circuit structure and task-driven learning, with partial symmetry breaking providing a low-dimensional control principle for balancing stability and flow in sequence generation.