TAM turns verified task advancement into online state and reusable operations for GUI agents
Submitted October 27, Task Advancement Memory (TAM) uses what has been reliably accomplished at two timescales: it maintains verified progress, remaining goals and visual evidence during a task, then extracts reusable operations from locally successful segments after the task, including segments inside an overall failed workflow.
Why it matters
Long GUI tasks require more than retaining an action transcript. An agent must distinguish established outcomes from missing inputs and preserve the productive parts of failed trajectories. TAM organizes memory around task results and operations with explicit verification conditions, giving cross-application continuity and failure-derived learning a clearer engineering boundary.
A closer look
During execution, the user goal is compiled into requirements and Advancement State updates completed requirements, object bindings and pending work from observed results; an Evidence Cache retains source-linked screenshots or crops when text is insufficient. At task end, a model-based reviewer considers local results plus later confirmation or contradiction, retaining supported operations with entry inputs, expected changes, verification conditions, object roles and GUI realizations. Under a shared Qwen3-VL-32B-Instruct backbone and 50-round budget on MobileWorld GUI-only, the base agent reaches 12.0% success, Advancement State alone 28.2%, and full TAM 34.2%; the strongest listed baseline, MemGUI-Agent, reaches 29.1%. TAM averages 20.6 steps and 29.2 EQA@50, and the authors report 38.7% fewer lifecycle tokens per solved task than the strongest baseline. On SPA-Bench English Level 3 with GPT-4o and Set-of-Mark, TAM reports 57.0% versus GUI-explorer's previously reported 53.7%.
Who this is for
Research and engineering teams building long-horizon GUI agents, cross-app workflows, task-state compression, visual evidence caches, experience memory and terminal verification
Original sources
Revisions & corrections (1)
— 新论文直接研究 GUI 代理如何把已核验进展用于在线执行与跨任务操作学习；日期、方法、比较条件、结果和限制均由 arXiv 记录与全文核验，适合原创双语研究摘要并并入任务终态/主动核验趋势。