MementoGUI: Learning Agentic Multimodal Memory Control for Long-Horizon GUI Agents

Researchers introduced MementoGUI, a plug-in memory framework for long-horizon GUI agents.

Its learned controller decides what to compress, store, and retrieve from multimodal interaction history.

By preserving key visual-textual evidence, it reduces context overload and improves decision-making over many steps.

The system outperformed history-replay and text-only baselines on benchmarks such as GUI-Odyssey, MM-Mind2Web, and MementoGUI-Bench.