Long-horizon GUI agents need to remember what they have seen to keep track of the task state. Keeping every historical screenshot is expensive, since each one adds visual tokens and slows inference. Writing textual summaries is cheaper, but those summaries can drop details that matter for later decisions.
ECHOES, described in a new arXiv preprint, takes a different route. Instead of storing screenshots or writing text, it compresses the visual history recursively into a compact latent representation. The agent is trained with privileged on-policy distillation: a teacher model that sees the full history guides the student, which only sees the compressed memory. This pushes the compressed representation to retain the information the agent actually needs.
The authors report that this reduces visual-token usage and inference costs while avoiding the state loss of textual summaries. Because this is a single preprint, there are no independent results or contrasting claims to compare against yet.