Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents
CarryOn reframes agent memory as reusable task objects, cutting token use while improving benchmark results.
The paper proposes an object-oriented context runtime for long-horizon agents, replacing repeated history processing with “continuation carriers,” execution context, and a live recent tail. Its update method revises only affected carrier records against evidence and environment checks, while preserving unaffected records across context folds. In experiments, CarryOn reports gains of 5.8% on OSWorld 2.0 and 7.0% on BrowseComp-Plus, with token reductions of 51.9% and 22.9%. Source: ArXiv · AI/CL/LG's note.
The paper proposes an object-oriented context runtime for long-horizon agents, replacing repeated history processing with “continuation carriers,” execution context, and a live recent tail. Its update method revises only affected carrier records against evidence and environment checks, while preserving unaffected records across context folds. In experiments, CarryOn reports gains of 5.8% on OSWorld 2.0 and 7.0% on BrowseComp-Plus, with token reductions of 51.9% and 22.9%. Source: ArXiv · AI/CL/LG's note.
score 5