Can sequence models learn the hidden memory in simplified dynamical systems?

When you simplify a complex dynamical system by removing variables, the remaining dynamics develop "memory"—they depend on past states, not just the present. This paper treats that memory as a sequence modeling problem, using Mamba (a fast state-space model) to predict the closure term from the system's resolved trajectory. The model trains efficiently on long sequences but runs in constant time per step during rollout. On Burgers' equation and chaotic Lorenz '96, it beats GRU-based and classical Markovian approaches at both accuracy and long-time stability.