Memory-augmented mobile GUI agents improve by storing successful execution trajectories and reusing them in later tasks. But in practice, a stored trajectory rarely matches a new task exactly. The new task may use different parameters, or share only a portion of the workflow, making direct reuse ineffective.
DeltaReplay, described in a new arXiv preprint, addresses this by shifting from exact trajectory matching to task-relative memory reuse. Rather than requiring a stored run to align perfectly with the current goal, the method appears to focus on the relationship between the stored task and the new one, allowing partial or parameter-shifted matches to still inform the agent's actions.
The abstract is truncated, so specific implementation details are not available from the source. This summary is based solely on the provided text; no independent evaluation or comparison with other methods is included.