Project Cairn:让编程智能体跨会话保留项目教训的开源 Skill
开源 Skill Project Cairn 用 AGENTS.md、日志与路线图把决策与坑点留在项目内,并在人工确认后才把可复用结论写入长期知识库,兼容 Claude Code 与 Codex 等。
编程智能体常在新会话里把项目里已经踩过的坑再踩一遍。开源项目 Project Cairn 以 Agent Skill 形式,让决策、坑点、进度与结论跟着项目走,并兼容 Claude Code、Codex 等读取项目规则的智能体。
本地结构包括:用 AGENTS.md 承载进入项目时智能体应遵守的规则;cairn/LOG.md 记录发生了什么(短摘要与指针);cairn/ROADMAP.md 汇总目标、计划与当前进度;cairn/.md 保存某一主题的当前结论。关键约束是「人工确认后才毕业」:智能体可以提出进入长期知识库的候选,但不能在无人确认时写入长期库。许可为 MIT。
对多会话、多工具切换的团队,这相当于给仓库加一层可检查的「项目记忆」;落地时仍要约定谁有权确认毕业、如何避免把密钥或客户数据写进可共享笔记,以及与现有 Obsidian/Notion 等库的同步边界。
综合来看,读者可先核验官方渠道与发布时间,再结合自身业务边界与合规要求决定是否跟进试点;本文仅整理公开信息,不构成采购、投资或法律建议,关键数字与能力边界请以一手文档为准。
Project Cairn Keeps Agent Lessons Across Sessions
Your agent shouldn’t have to rediscover a project’s hard-won lessons.
Project Cairn is an AI-agent skill for builders who want project knowledge to survive across sessions and compatible agents such as Claude Code and Codex.
It helps you keep decisions, pitfalls, progress, and conclusions connected to the project by storing rules and working context locally, then preparing reusable lessons for a knowledge base only after human confirmation.
Key features:
• Project rules – uses AGENTS.md to carry the rules an agent reads when it enters the project • Working history – cairn/LOG.md records what happened with short summaries and pointers • Current direction – cairn/ROADMAP.md keeps the overall goal, plan, and current progress together • Scoped conclusions – cairn/.md holds the current conclusion for one subject • Human-confirmed reuse – the agent can propose a graduation candidate, but cannot write to the long-term knowledge base without confirmation
It’s open-source (MIT license).
Link in the reply 👇