Multi-session LLM agents—systems that maintain a conversation history across interactions—sometimes switch underlying models to save cost or improve performance. That switch requires passing the prior history to the new model, but the consequences of that transfer are not well understood.

A new arXiv preprint, "New Model, Old Habits: Behavior Transfer under Model Swapping in Multi-Session LLM Agents," takes up this question. The authors note that while model swapping is common, its effects on the recipient model remain underexplored. They conduct controlled experiments to examine how behavior transfers when history is moved from one LLM to another.

The abstract frames the work as a step toward understanding continuity in agent behavior across model changes. Specific results are not included in the announcement, so the paper's conclusions will need to be read in the full manuscript. The title hints that old behavioral patterns may persist after a swap.