Motor Learning vs Episodic Memory in Neural Networks: Why Hebbian Plasticity Succeeds at Calibration but Fails at Recall

The mammalian brain segregates learning into distinct systems: cortical networks excel at motor skills and habits, while hippocampal circuits specialize in episodic memory18-20. This architectural division has puzzled neuroscientists for decades—why not use one flexible learning mechanism for everything? Here we show that these systems are not just specialized, but computationally incompatible. Using recurrent neural networks with Hebbian plasticity in navigation tasks, we systematically tested whether fast synaptic modifications could support both motor learning and episodic memory. Hebbian plasticity dramatically improved motor performance—increasing target zone occupancy by 18% while reducing behavioral variance by 38%—but catastrophically impaired episodic recall, with memory performance dropping from 100% to 10% after plastic updates. This failure persisted despite forced memory encounters, rich spatial sensing, surprise-gated plasticity, and replay mechanisms. The dissociation arises because motor learning requires variance collapse to stabilize trajectories, while episodic memory demands variance preservation to separate events. These opposing computational pressures explain why evolution segregated learning into specialized circuits rather than relying on a unified mechanism. Our findings suggest that artificial intelligence systems attempting to use single learning mechanisms for multiple cognitive functions may face similar fundamental limitations, providing a principled basis for designing multi-system architectures that respect these computational constraints.