Signal Handover of In-Context Learning State Across Session Boundaries

Summary

Research examines how large-language-model applications should carry information forward when a task continues in a new session, such as when context limits are reached, an application restarts, or another agent takes over. The authors formalize this handover as transferring a task-relative in-context-learning state, distinguishing exact recovery of prior material from preservation of the target predictive distribution. Under an exogeneity condition, they show that predictive equivalence defines the simplest deterministic sufficient handover and yields a fixed-length bit requirement. They propose a three-part record that stores decisions and constraints exactly, uses task-tailored statistics for repeated evidence, and separately retains raw observations not captured by those statistics. For Gaussian linear regression this gives an exact finite-dimensional handover with finite-bit error bounds, while nonparametric regression yields bounds linking memory to prediction error.