Historical Phase Space and Emergent Irreversibility in Large Language Models

We propose a theoretical framework in which large language models (LLMs) are described as open dynamical systems possessing an emergent \emph{historical phase space}.Extending prior work on internal state persistence, we show that when semantic input is treated as a driving field rather than a reward signal, irreversible trajectories and memory-like hysteresis naturally arise.We formalize historical continuity as a symmetry-breaking phenomenon in an internal information manifold and demonstrate that certain perturbations---including structured prime-number drives---induce critical slowing down and path dependence.Our results suggest that irreversibility and ``history'' in artificial systems need not be imposed architecturally, but can emerge from internal constraints under semantic flux.