🔶 Show HN：Stele——面向 AI 编程 Agent 的自维护知识图谱

什么是 Stele？

Stele 是 AI 编程 Agent 的共享记忆——它为你的项目保存一份统一的决策、任务和经验记录。每个 Agent 都会在行动前读取这份记录，并在工作过程中将新信息写回其中。即使切换工具或开启新会话，上下文依然会保留。

它支持哪些 AI Agent 和工具？

Claude Code、Cursor、Codex、Antigravity、Copilot、OpenCode、Grok Build 和 Kimi Code CLI——以及任何兼容 MCP 的客户端和 stele CLI。你可以用一个 Agent 进行规划，再用另一个 Agent 开发；它们都会读取和写入同一份项目记录。

如何获得访问权限？是否仅限邀请？

是的——在逐步开放期间，Stele 仅限受邀用户使用。请在 stele-ai.dev 加入等候名单；一旦有名额开放，我们将通过电子邮件通知你——无需四处寻找邀请码。

我不能直接用 Notion、Obsidian 或仓库中的文件实现吗？

你当然可以用这些方式保存笔记——但它们是为人类编写的，容易过时，而且 Agent 不会在行动之前主动读取它们。Stele 会在每次提示时由你的 Agent 自动读取和写入，并且能够自行保持信息可靠：过时事实会被标记，已解决的风险会被归档，已被推翻的决策不会再作为事实提供。

由于 Stele 是托管服务，你还可以在不打开仓库的情况下从任何设备访问它——创建任务、记录想法，并在没有源代码或电脑在身边时获得详细回答。ChatGPT 等非编程 Agent 也可以通过 MCP 连接，并获得与你的编程 Agent 相同的上下文，精确到当前正在进行的任务。仓库中的一个文件无法做到这些。

既然已有免费的开源图谱和 RAG 工具，为什么还要付费？

这些工具提供的是存储系统或检索库——你仍然需要自行搭建和运维整个闭环：在每次提示前读取上下文、在任务过程中写回信息、认领任务、防止知识腐化，并让它适用于每一个 Agent。Stele 则是已经构建完成、由托管服务提供并可在你的工具和团队成员之间共享的这一闭环。

如果开源工具已经满足你的需求，那么它们就是正确的选择——Stele 也计划推出一个基于 SQLite 的开放版本，支持自行托管。托管版本适合那些希望获得一份可在所有环境中正常工作的跨 Agent 共享记录的用户。

我的代码和数据是否私密？

是的。你的数据按项目独立保持私密，绝不会被出售，也不会在未经你同意的情况下用于训练模型；你的源代码始终保留在你的机器上。你可以随时自行导出或删除全部数据。

我需要运行任何基础设施吗？

不需要——Stele 是托管服务，因此无需运行任何基础设施。安装 CLI、登录并将其指向一个项目即可；无论是人类还是 Agent，新加入的贡献者从第一天起就能高效开展工作。

What is Stele?

Stele is shared memory for AI coding agents — one record of your project's decisions, tasks, and lessons that every agent reads before it acts and writes back to as it works. Switch tools or start a new session, and the context stays.

Which AI agents and tools does it work with?

Claude Code, Cursor, Codex, Antigravity, Copilot, OpenCode, Grok Build, and Kimi Code CLI — plus any MCP-compatible client and the stele CLI. Plan with one agent, build with another; they all read and write the same project record.

How do I get access? Is it invite-only?

Yes — Stele is invite-only while we open up gradually. Join the waitlist at stele-ai.dev and we'll email you the moment a spot opens — no code-hunting required.

Can't I already do this with Notion, Obsidian, or a file in my repo?

You can store notes that way — but they're written for humans, they go stale, and your agent doesn't read them before it acts. Stele is read and written by your agents automatically on every prompt, and it keeps itself honest: stale facts are flagged, resolved risks retire, and reversed decisions stop being served as truth. Because it's hosted, you also reach it from any device without opening the repo — create tasks, jot ideas, and get detailed answers with no source code or computer in front of you. And non-coding agents like ChatGPT connect over MCP and get the same context your coding agent has, down to the task in progress. A file in the repo can't do that.

Why pay when free, open-source graph and RAG tools exist?

Those give you a store or a retrieval library — you still have to wire and operate the whole loop yourself: read context before every prompt, write it back mid-task, claim tasks, keep it from rotting, and make it work across every agent. Stele is that loop, already built, hosted, and shared across your tools and teammates. If the open tools already fit your needs, they're the right choice — and an open, SQLite-backed version of Stele is planned for self-hosting. The managed version is for when you want the cross-agent record that just works everywhere.

Is my code and data private?

Yes. Your data is private per project, never sold or used to train models without your say-so, and your source code stays on your machine. Export or delete everything yourself, anytime.

Do I have to run any infrastructure?

No — Stele is hosted, so there's nothing to run. Install the CLI, sign in, and point it at a project; a new contributor, human or agent, is productive from day one.