The paper challenges the common practice of treating a whole trajectory as the unit of experience when improving GUI agents. Instead, it splits each trajectory into four components: locators, procedures, state facts, and lessons. Each component is then sent either to the model's weights via fine-tuning or to the context via retrieval, and the two destinations are compared on identical items across three backbone families and two environments.
The key finding is that components differ in where they work best: locators and lessons gain more from being put in the weights, while procedures and state facts benefit more from being kept in the context. A simple rule based on two properties measurable before training, recurrence and state-conditionality, correctly predicts the best destination for held-out backbone families in 24 out of 24 cases. Routing by this rule beats the better single-destination approach by 3.5 points on average, and the reverse assignment hurts by 7.8 points. This explains why prior whole-trajectory comparisons disagreed: the mix of components in a trajectory determines which destination appears superior.