AI News AI资讯 7d ago Updated 7d ago 更新于 7天前 48

OpenAI's Computer History turns your clicks and keystrokes into a searchable ChatGPT memory timeline OpenAI的Computer History将你的点击和按键操作转化为可搜索的ChatGPT记忆时间线

OpenAI's Computer History replaces the screenshot-based "Chronicle" preview with an event-tracking system that records clicks, keystrokes, shortcuts, and app switches via macOS accessibility APIs Captured interaction events are converted into plaintext Markdown memory files stored locally, which ChatGPT and Codex can use as contextual memory The feature detects recurring workflows and suggests reusable "Skills" that Codex can turn into automation templates Memory files are unencrypted plaintext OpenAI推出Computer History功能,通过macOS辅助功能系统记录用户点击、键盘输入和应用切换等交互事件,生成可搜索的时间线记忆供ChatGPT和Codex使用 该功能替代了依赖截图的"Chronicle"预览版,改为记录结构化交互事件,不捕获屏幕内容、麦克风或系统音频,隐私保护更完善 系统可识别重复工作流并建议生成可复用"技能",用户可通过Codex将记录的工作流转化为自动化模板 记忆文件以明文Markdown格式本地存储且未加密,存在同账户程序读取风险,同时OpenAI警告存在提示注入攻击风险 功能采用多层权限控制(企业管理员+用户双重授权),暂不支持欧洲经济区、瑞士和英

72
Hot 热度
65
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • OpenAI's Computer History replaces the screenshot-based "Chronicle" preview with an event-tracking system that records clicks, keystrokes, shortcuts, and app switches via macOS accessibility APIs
  • Captured interaction events are converted into plaintext Markdown memory files stored locally, which ChatGPT and Codex can use as contextual memory
  • The feature detects recurring workflows and suggests reusable "Skills" that Codex can turn into automation templates
  • Memory files are unencrypted plaintext stored locally; temporary event files are deleted after 48 hours and processed server-side without retention
  • OpenAI warns of heightened prompt injection risks and advises against using the feature with sensitive communication or health/financial apps without explicit consent

Why It Matters

Computer History represents a significant shift toward persistent, behavior-aware AI assistants that can recall and act on a user's actual workflow patterns rather than relying on one-off prompts. The move from screenshots to structured event tracking raises important questions about privacy, data security, and the trust model between users and AI platforms. For AI practitioners, it signals the growing industry trend of embedding AI deeper into desktop workflows through local-first memory systems.

Technical Details

  • Event capture mechanism: Uses macOS accessibility APIs to record interaction events (clicks, keystrokes, keyboard shortcuts, app switches) rather than visual content like screenshots or screen recordings; excludes private browsing mode entirely
  • Memory pipeline: Raw event files are stored temporarily in the ChatGPT app group and deleted after 48 hours; they are then processed server-side to generate plaintext Markdown memory summaries, which are written locally and retained until manually deleted by the user
  • Skill detection and automation: ChatGPT analyzes the timeline for recurring workflow patterns and surfaces them as suggested "Skills"; users can then prompt Codex to generate reusable automation templates from those recorded patterns
  • Data handling: Memory files are unencrypted plaintext on the local filesystem, readable by any process running under the same macOS user account; OpenAI states the files are not used for AI training, though chat content containing memories may become training data depending on user data control settings
  • Permission architecture: Multi-layered opt-in system requiring both workspace admin approval (Business/Enterprise) and individual user consent; users maintain granular control via include/exclude lists for apps and websites, with pause/resume capability from the menu bar

Industry Insight

  • The shift from screenshot-based to event-based tracking reflects a broader industry pivot toward privacy-preserving AI memory systems; competitors will likely follow suit, making structured interaction logging a standard pattern for desktop AI assistants
  • The unencrypted local storage and prompt injection warnings highlight an emerging class of security risks for AI-augmented productivity tools—developers building similar features should prioritize encryption, sandboxing, and injection-resistant context pipelines from day one
  • The multi-layered consent model (admin + user + per-app controls) sets a precedent for enterprise AI deployments, suggesting that organizational governance will become a critical differentiator as AI memory features proliferate across the workplace

TL;DR

  • OpenAI推出Computer History功能,通过macOS辅助功能系统记录用户点击、键盘输入和应用切换等交互事件,生成可搜索的时间线记忆供ChatGPT和Codex使用
  • 该功能替代了依赖截图的"Chronicle"预览版,改为记录结构化交互事件,不捕获屏幕内容、麦克风或系统音频,隐私保护更完善
  • 系统可识别重复工作流并建议生成可复用"技能",用户可通过Codex将记录的工作流转化为自动化模板
  • 记忆文件以明文Markdown格式本地存储且未加密,存在同账户程序读取风险,同时OpenAI警告存在提示注入攻击风险
  • 功能采用多层权限控制(企业管理员+用户双重授权),暂不支持欧洲经济区、瑞士和英国地区

为什么值得看

Computer History代表了AI助手从被动响应向主动记忆演进的标志性产品方向,通过本地活动追踪构建个性化上下文,为AI Agent的长期记忆能力提供了可行的产品化路径。其隐私保护设计与安全风险警示的并存,为行业在个人数据收集与AI能力扩展之间寻找平衡点提供了重要参考案例。

技术解析

  • 数据采集架构:通过macOS Accessibility API捕获用户交互事件(点击、键盘输入、快捷键、应用切换),替代了早期Chronicle的截图方案,避免捕获屏幕内容、麦克风输入和系统音频,隐私边界更清晰
  • 本地存储与处理流程:临时事件文件存储在ChatGPT应用组内,48小时后自动删除;处理后生成的记忆文件以明文Markdown格式保留在本地文件系统,需用户手动删除;记忆生成在OpenAI服务器端完成,处理后不保留原始事件文件
  • 智能工作流识别:系统自动检测用户活动中的重复模式,当ChatGPT识别出可复用工作流时,会在时间线中建议创建"Skill",用户可进一步要求Codex基于记录生成自动化模板
  • 权限与安全机制:企业/商业工作区需管理员显式启用后,各用户仍需单独授权并开启Memories功能;用户可通过包含/排除列表控制数据源应用,支持从菜单栏随时暂停/恢复录制
  • 风险提示设计:OpenAI文档明确标注提示注入风险——网页中嵌入的恶意指令可能通过记录的活动内容影响ChatGPT/Codex行为;建议对涉及健康、财务等敏感数据的应用暂停录制或排除

行业启示

  • AI Agent记忆能力产品化加速:Computer History标志着AI助手从"无状态对话"向"有记忆交互"的关键转变,未来个人AI助手将依赖此类长期记忆系统提供个性化、连续性的服务体验
  • 隐私合规成为AI功能上线的先决条件:OpenAI在功能设计中内置多层授权、地域限制和明确的风险警示,反映了AI产品在全球不同监管环境下的合规策略,企业需在功能创新与隐私保护之间建立系统化框架
  • 本地优先的数据处理模式将成为趋势:通过本地事件捕获、临时文件处理和用户控制的数据排除机制,OpenAI展示了"边缘采集+云端处理"的混合架构,这种模式在保障用户体验的同时降低数据泄露风险,为AI数据收集提供了新的设计范式

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPT GPT Closed Source 闭源 Product Launch 产品发布 LLM 大模型 Agent Agent