Endogenous Landscape Dynamics for Continual World Models

Continual world models must absorb new experience without either erasing useful structure or becoming unable to adapt. Most approaches address this stability--plasticity problem through parameter updates, replay, regularization, or explicit memory. We propose a complementary object of adaptation: the geometry that governs the model's own trajectories. Endogenous Landscape Dynamics (ELD) couples a fast interaction trajectory to a slowly evolving landscape that is locally reshaped by experience and gradually relaxes toward a prior geometry. This minimal construction makes memory, forgetting, interference, and recovery properties of one dynamical system. It also yields a distinctive prediction: agents presented with the same current input can respond differently because their interaction histories have constructed different landscapes. We formulate ELD as an idea for continual world models, identify its stability--plasticity controls, and propose an experimental protocol that can falsify the account by separating beneficial history dependence from indiscriminate hysteresis.