Our work provides a mechanistic model of motor memory formation that: Demonstrates savings from intrinsic dynamics, supports a causal role for persistent neural traces, and bridges recent neurophysiological findings.

We perturb the RNN’s preparatory activity along and against the neural shift direction. Perturbing with the shift enhances savings, while perturbing against it reduces or eliminates savings. This suggests a causal link between the trace and savings.

Our model predicts in silico the same neural trace of prior learning seen in vivo in monkey motor cortex (Sun, O’Shea et al., Nature 2022): A shift in preparatory activity that persists after washout, without affecting movement.

Savings refers to faster relearning after prior exposure to a motor task. We show that an RNN trained to control a realistic model of the arm predicts savings in motor learning without any explicit cues. This suggests a neural population-level account of how savings can emerge implicitly.

In this work, we use a computational model to test the idea that savings in motor learning can emerge from neural population dynamics—even without explicit cues or cognitive strategies.