Mobile GUI agents often keep successful execution trajectories in a memory bank and try to reuse them for later tasks. In practice, a stored trajectory rarely matches a new task exactly. The new task may use different parameters, share only some steps, or have no relevant record at all. Forcing the agent to follow irrelevant memory can mislead it, while discarding potentially useful memory deprives it of past guidance.
To tackle this dilemma, we introduce DeltaReplay – a step‑level memory reuse framework that decides how to use existing records without altering them. We observe that the reusable portion of a trajectory is determined not by the record itself but by its relation to the new task, mainly through page‑level consistency and action‑level generality.
Accordingly, we store trajectories as paths in a transition graph where nodes are pages and edges are actions between pages. Each edge’s action is split into a task‑independent operation and task‑specific parameters. At reuse time, DeltaReplay compares each recorded step with the current screen and the new task, then decides whether to follow the step as‑is, execute it after substituting parameters, or leave it to the base agent.
Experiments on AndroidWorld and SPA‑Bench show that, with the same backbone model, DeltaReplay improves task success rates by up to 10.3 and 25.0 percentage points respectively. These results demonstrate that making per‑step decisions about memory usage enables agents to benefit even from partially matching trajectories.
Review: DeltaReplay’s graph‑based representation and parameter separation provide fine‑grained memory reuse, markedly boosting the robustness and efficiency of mobile GUI agents and offering a fresh direction for memory‑augmented intelligent systems.