Handover of In-Context Learning State Across Session Boundaries
Agent handoffs should preserve task-relevant predictive state, not blindly compress the transcript. The paper offers a three-part record for carrying decisions, evidence, and irreducible observations across sessions.
The paper frames session handover as transferring a task-relative in-context learning state. Its **three-part record** keeps decisions and constraints exactly, summarizes repeated evidence with task-justified statistics, and retains observations whose effects those statistics lose.
For coding-agent workflows, design handoff files around the continuation task rather than generic transcript summaries. Preserve hard constraints verbatim, compress only evidence with a known sufficient representation, and keep raw observations when their downstream effect is uncertain.
The paper frames session handover as transferring a task-relative in-context learning state. Its **three-part record** keeps decisions and constraints exactly, summarizes repeated evidence with task-justified statistics, and retains observations whose effects those statistics lose. For coding-agent workflows, design handoff files around the continuation task rather than generic transcript summaries. Preserve hard constraints verbatim, compress only evidence with a known sufficient representation, and keep raw observations when their downstream effect is uncertain. The sufficiency result assumes **exogeneity**, while exact finite-dimensional handover is demonstrated for Gaussian linear regression. The nonparametric case provides memory-versus-error bounds, not a universal recipe for software tasks.
This gives durable agent state a task-relative compression rule: preserve decisions and constraints exactly, summarize only evidence with a justified sufficient representation, and retain raw observations when effects are uncertain. It strengthens reset and shared-plan patterns while narrowing them with formal assumptions that do not establish a universal coding-agent handoff format.