ECHOES Agent Solves Memory Problem in GUI Automation

A GUI agent that has to click through forty screens to book a flight or fill out a form runs into a basic math problem: keep every screenshot and the token bill explodes, throw them away and replace them with text summaries and the agent starts forgetting what it already clicked. Researchers behind a new system called ECHOES are tackling that trade-off directly, framing it as a question of what long-horizon agents actually need to remember versus what they can afford to carry around.

The work, described in a paper on ECHOES and a companion benchmark called ECHOESBMK, targets a failure mode familiar to anyone who has watched an automated agent loop endlessly on a task it already completed, because its compressed memory dropped the detail that mattered. The authors built their benchmark from real GUI agent failure trajectories specifically to separate perception errors from memory errors, testing both at once.

Across four existing GUI benchmarks plus their own diagnostic set, the team reports gains in task completion and in what they call memory fidelity, without needing to store screenshots during deployment. That distinction, performance without the usual visual-history overhead, is the core claim the rest of the paper builds out in detail.