MementoGUI
MementoGUI是一种面向长周期GUI智能体的多模态记忆管理框架，旨在增强自动化图形界面操作任务中的跨步骤记忆保持与调用能力。该框架采用插件式架构设计，以预训练视觉语言模型为主干，通过轻量级适配器模块实现记忆的编码、存储与检索控制，适用于需要持续感知与推理的复杂GUI自动化场景。
All Facts (6)
MementoGUI的记忆控制器（Memento Core）基于共享的冻结QwenVL主干，配备四个任务特定的LoRA适配器
95%UnverifiedMementoGUI的训练数据自动标注管道中，研究人员随机抽取200条轨迹做人工验证，197条被判定完全正确
90%UnverifiedMementoGUI的训练数据来自PSAI计算机使用轨迹的自动生成，无需大量人工标注
85%UnverifiedThe MementoGUI framework is designed with a dual-timescale memory system comprising two complementary memory levels: working memory and episodic memory
60%UnverifiedMementoGUI adopts a plug-in architecture that does not require fine-tuning the original GUI action model, enabling zero-fine-tuning plug-and-play integration
60%UnverifiedMementoGUI reframes the long-horizon GUI agent amnesia problem as a multimodal memory management problem rather than a simple context length problem
60%