CONTEXT-DEPENDENT PHASE RECONFIGURATION SEPARATES CONFIGURATION FROM INTRINSIC COMPUTATION

Recurrent systems often need to change their context-dependent representations while preserving useful computations. We study this problem in a periodic continuous-attractor model of hippocampal-to-entorhinal reconfiguration. A compatibility condition separates environment-dependent phase offsets from the spatial derivative that defines path integration. Structure-preserving phase translation follows a state-dependent tangent direction whose family lies in a fixed two-dimensional first-harmonic subspace per module. With recurrent dynamics frozen, a learned high-dimensional feedback interface expresses this predicted effective control geometry. Paired projection interventions support the functional role of the local tangent component, while post-capture interventions show that compatible feedback can remain part of ongoing closed-loop computation. An independent recurrent system dissociates target representation, instantaneous latent phase, and phase dynamics. In an external waypoint task, preserving dynamics provides a growing behavioral advantage after reconfiguration; this advantage persists after matching latent phase and external kinematics. Together, these results frame adaptation as computation-preserving reconfiguration, evaluated through both the state reached and its subsequent evolution.