Endogenous Landscape Dynamics for Continual World Models

Continual world models face a representational problem: where should experience persist, how strongly, and how should that state constrain future interaction? Most approaches address stability and plasticity through parameter updates, replay, regularization, or explicit memory. We propose a complementary object of adaptation: the geometry governing the model's trajectories. Endogenous Landscape Dynamics (ELD) couples a fast interaction trajectory to a slowly evolving landscape that experience locally reshapes and that relaxes toward a prior geometry. Deposition is gated by consequence magnitude, so a rare high-impact event allocates pronounced local structure while routine traversal leaves shallow traces, and valence sets its sign, making appetitive memory attracting and aversive memory deflecting. This places allocation, forgetting, interference, and recovery in one system and predicts different responses to identical inputs after different histories. ELD does not resolve the stability--plasticity trade-off or supply a complete architecture; it isolates an explicit, causally testable state and a protocol that distinguishes useful geometric memory from indiscriminate hysteresis.