- 昨日,Anthropic 的 Claude Code 源代码意外泄露,涉及超过 51.2 万行代码和 2000 多个文件。分析发现,除现有功能外,代码中包含多个被禁用或隐藏的功能,揭示了公司未来可能的开发方向。其中最受关注的是名为 Kairos 的后台常驻守护进程,可在终端关闭后持续运行,通过周期性“”提示检查是否需要执行新操作,并支持“PROACTIVE”标志以主动向用户推送未请求但重要的信息。
Kairos 依赖一个基于文件的“记忆系统”,旨在跨用户会话保持状态。代码中的隐藏提示表明,该系统目标是全面记录用户身份、协作偏好、应避免或重复的行为以及任务背景。为整合记忆,代码还提及名为 AutoDream 的系统,当用户空闲或手动休眠时,会触发“梦境”过程,引导 AI 回顾当日交互记录,识别需持久化的新信息,合并内容以避免重复与矛盾,并清理冗余或过时记忆,同时监测“记忆漂移”问题。
Kairos 实现后台持续运行
记忆系统支持跨会话状态保持
AutoDream 模拟 AI 自我反思过程
主动推送功能提升交互前瞻性
- The source code leak of Anthropic’s Claude Code revealed over 512,000 lines of code across more than 2,000 files, exposing both active infrastructure and disabled or hidden features that hint at future development directions. Among the most significant findings is "Kairos," a persistent background daemon designed to operate even when the terminal interface is closed. Kairos uses periodic "" prompts to assess whether new actions are required and includes a "PROACTIVE" flag intended to surface relevant information without explicit user requests. The system relies on a file-based memory architecture that persists across user sessions, aiming to maintain a comprehensive understanding of user preferences, collaboration styles, and contextual work patterns.
Additionally, the code references an "AutoDream" system, activated during user idle time or manual sleep commands. AutoDream instructs the AI to perform a reflective review of memory files, identifying new information to retain, consolidating data to avoid redundancy or contradictions, and pruning outdated or verbose memories. It also monitors for "drifted" memories—inconsistencies that have been observed in user-implemented memory systems. These features suggest a long-term vision for autonomous, context-aware AI assistants with continuous learning capabilities.
Key Takeaways:
Kairos enables background operation and proactive task management
AutoDream supports memory consolidation and drift detection across sessions
Claude Code aims for persistent, personalized user context
Leaked code reveals roadmap for autonomous AI behavior
Source: Original Article
查看原文 →
View Original →