Intel / article

MemU-Bench: Benchmarking GUI Agents on Long-Horizon User Interactions

Source: arxiv.org
Author: Yizhou Liu, Zhengxi Lu, Fei Tang, Yiwen Qiu, Yong Du, Tongbo Chen, Bofan Chen, Tao Jiang, Wenhao Xu, Yunfeng Shu, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
Date: Published Nov 3, 2026

Context

MemU-Bench tests whether phone-using AI agents can act on vague requests that depend on what the user did earlier, the way people actually ask for things (arXiv). Most GUI- benchmarks give precise, self-contained instructions. Here, 118 tasks are built on 29 long recorded user sessions, and the agent has to turn that raw interaction history into memory and pull out the missing detail when a request leaves it out. Results are checked against the apps' saved databases. Besides success rate, the scores how efficiently the agent got there, so an agent that succeeds by searching through every app is not credited as if it remembered.

According to the abstract, without the user's history even the strongest system tested, Gemini-3.1-Pro, reached only 16.4% on the path-length-weighted success score, and end-to-end GUI agents fell below 5%. The authors also propose AppMem, a method that needs no training: it organizes the history into memory per app, tracks tasks the user abandoned when switching apps, and retrieves memories by first guessing which app the request concerns. They report it lifted GPT-5.5 on that score from 15.7% to 66.9% and beat a plain summary of the history by 9.3%. These are the authors' own results on their own benchmark.

Key takeaways · AI-distilled

MemU-Bench contains 118 vague mobile tasks in 29 long-horizon user trajectories, with task outcomes verified against serialized application databases.
Without user history, the best system, Gemini-3.1-Pro, reached only 16.4% Success weighted by Path Length (SPL), and end-to-end GUI agents fell below 5%, according to the authors.
The proposed AppMem framework is training-free: it distills history into an app-centric memory bank, tracks intents interrupted by app switches, and retrieves memories by inferring the target app.
AppMem raised GPT-5.5's SPL from 15.7% to 66.9% and beat a naive history-summary baseline by 9.3%, with consistent gains across the backbones evaluated.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.

Read the source arxiv.org