完整的文档索引见llms.txt。在网址后加.md可以拿到 Markdown 版本；本页的 Markdown 版本是standalone-tree.zh-CN.md。

指南

仅安装世界树服务

不部署完整的 OpenMousse，单独运行 `mousse-tree` 为外部 AI 应用同步记忆。

完成本指南后，你的服务器上将单独运行轻量级的世界树服务（mousse-tree），为 Claude、ChatGPT、Gemini、Notion、Claude Code 及 OpenClaw 提供共用的长期记忆。本指南适用于不需要部署完整 OpenMousse 客户端与多 Agent 系统、仅需跨平台统一记忆的用户。

前提条件

•一台具备 Python 3.11 以上版本与 git 的 Linux、macOS 或 Windows WSL 2 机器。
•若需接入 Claude 网页版、ChatGPT、Gemini 或 Notion，机器需具备公网 HTTPS 访问能力（推荐使用免费的 Tailscale Funnel）。

隐私说明

•不调用模型、不产生额外模型费用：mousse-tree 仅负责记忆的存储、全文检索（FTS5 trigram）与 MCP 接口响应，本身不发起任何 LLM 调用；各平台的对话仍使用你原有的订阅。
•数据留在本地：所有记忆保存在你服务器的 ~/.mousse-tree/tree.db（SQLite）或你指定的本地 Markdown 文件夹中。软链接与隐藏文件不会被收录，MCP 工具无法访问笔记文件夹以外的路径。
•可审计的遗忘机制：调用 forget 删除记忆时，系统会清空正文、将空壳移入归档并删除检索索引，仅保留不含正文的 ID 与操作时间记录；若笔记文件夹开启了第三方云同步（如 iCloud、Obsidian Sync 或 git），其历史版本中可能仍留存旧内容，需在对应同步工具中清理。

步骤

1. 安装并初始化常驻服务

在服务器终端依次运行以下命令（将称呼与时区替换为你的实际信息）：

终端

pipx install "git+https://github.com/openmousse/openmousse#subdirectory=tree"mousse-tree init --name 你的称呼 --tz UTC --lang zhmousse-tree install-service

•第一行也可使用 pip install --user "git+https://github.com/openmousse/openmousse#subdirectory=tree"。
•install-service 会配置开机常驻服务（Linux 使用 systemd，macOS 使用 launchd），默认仅监听本机 127.0.0.1:8787。
•若本机同时运行 OpenClaw，可运行 mousse-tree install-openclaw && systemctl --user restart openclaw-gateway，将世界树导出目录加入 OpenClaw 的检索路径。

2. 选择存储方式（可选切换为 Markdown 笔记库）

服务默认使用 SQLite 存储（~/.mousse-tree/tree.db）。如果你希望将每条记忆存为一篇独立的 Markdown 笔记（例如放入 Obsidian 笔记库以便在手机上查看和编辑），可运行迁移命令：

终端

mousse-tree migrate markdown --notes ~/vault/世界树 --profile-note 档案.mdsystemctl --user restart mousse-tree

•迁移后，笔记文件成为数据源，SQLite（index.db）仅作为检索索引；你在笔记软件中直接新建或修改笔记后，服务会在 5 秒内自动同步索引。
•被新记忆取代（supersedes）或你手动移入 归档/（英文为 Archive/）子目录的笔记不再参与检索。
•如需切回 SQLite 存储，运行 mousse-tree migrate sqlite 并重启服务即可。

3. 使用 Tailscale Funnel 开放公网路径

云端 AI 平台需要通过公网 HTTPS 访问世界树。使用 Tailscale Funnel 仅对外暴露 /t 与 /m 两个 MCP 路径前缀：

终端

tailscale funnel --bg --set-path=/t http://127.0.0.1:8787/ttailscale funnel --bg --set-path=/m http://127.0.0.1:8787/mmousse-tree init --host <你的机器名>.<tailnet>.ts.netsystemctl --user restart mousse-tree

注意：请勿公开管理页

切勿使用 Funnel 暴露 /ui 管理页路径。管理页（含记忆浏览、修改、待确认审批与档案编辑）应仅通过本机、tailscale serve（仅限局域网内设备）或 SSH 隧道访问。如需让外部平台写入的内容先进入「待确认」，可在 ~/.mousse-tree/config.json 中设置 "require_confirm": true。

4. 查看接入地址并连接各平台

在服务器终端运行以下命令，打印各平台的专属接入地址与本地管理页链接：

终端

mousse-tree urls

复制对应平台的地址，按照指南「把你常用的 AI 接进世界树」第 2、3 步的表格说明，将地址与自定义指令填入 Claude、ChatGPT、Gemini、Notion 或 Claude Code 中。

成功的标志

当以下各项均成立时，说明独立世界树服务已正常运行：

下一步

•把你常用的 AI 接进世界树：查看各平台连接器的具体配置路径与自定义指令原文。
•世界树：了解世界树的设计细节与五类记忆来源。

最后更新于 2026-10-19这一页有帮助吗？

For a complete index of the docs, see llms.txt. Append .md to any URL for the Markdown version; this page's Markdown is at standalone-tree.md.

Guides

Install the Standalone Memory Tree

Run `mousse-tree` on its own to sync memory across external AI apps without deploying the full system.

After completing this guide, your server will run the lightweight standalone memory tree service (mousse-tree), providing a shared long-term memory for Claude, ChatGPT, Gemini, Notion, Claude Code, and OpenClaw. This guide is for users who want unified cross-platform memory without deploying the full OpenMousse client and multi-agent system.

Prerequisites

•A Linux, macOS, or Windows WSL 2 machine with Python 3.11 or later and git.
•Public HTTPS access to the machine if you want to connect Claude on the web, ChatGPT, Gemini, or Notion (free Tailscale Funnel is recommended).

Privacy

•No model calls or extra model costs: mousse-tree only handles memory storage, full-text search (FTS5 trigram), and MCP endpoint responses. It makes no LLM calls of its own; conversations on each platform continue to use your existing subscriptions.
•Data stays local: All memories are stored on your server in ~/.mousse-tree/tree.db (SQLite) or in a local Markdown folder you specify. Symlinks and hidden files are ignored, and MCP tools cannot access paths outside the notes folder.
•Auditable forgetting: Calling forget to delete a memory clears the body, moves the empty shell to the archive, and deletes the search index entry, retaining only the bodyless ID and timestamp record. If third-party cloud sync (such as iCloud, Obsidian Sync, or git) is enabled on your notes folder, older versions may still retain the previous content and must be purged in that sync tool.

Steps

1. Install and initialize the background service

Run the following commands in order on your server terminal (replace the name and time zone with your own):

Terminal

pipx install "git+https://github.com/openmousse/openmousse#subdirectory=tree"mousse-tree init --name YourName --tz UTC --lang zhmousse-tree install-service

•For the first line, you can also use pip install --user "git+https://github.com/openmousse/openmousse#subdirectory=tree".
•install-service configures a background service that starts on boot (systemd on Linux, launchd on macOS) and listens only on 127.0.0.1:8787 by default.
•If OpenClaw is running on the same machine, run mousse-tree install-openclaw && systemctl --user restart openclaw-gateway to add the memory tree export directory to OpenClaw's search path.

2. Choose a storage backend (optionally switch to a Markdown vault)

By default, the service stores data in SQLite (~/.mousse-tree/tree.db). If you want each memory stored as a separate Markdown note (for example, inside an Obsidian vault so you can view and edit notes on your phone), run the migration command:

Terminal

mousse-tree migrate markdown --notes ~/vault/世界树 --profile-note 档案.mdsystemctl --user restart mousse-tree

•After migration, the note files become the source of truth and SQLite (index.db) serves only as the search index. When you create or edit notes directly in your note-taking app, the service syncs the index automatically within 5 seconds.
•Notes replaced by a newer memory (supersedes) or moved manually into the 归档/ (Archive/ in English) subdirectory are excluded from search.
•To switch back to SQLite storage, run mousse-tree migrate sqlite and restart the service.

3. Expose public paths with Tailscale Funnel

Cloud AI platforms need public HTTPS access to reach the memory tree. Use Tailscale Funnel to expose only the /t and /m MCP path prefixes:

Terminal

tailscale funnel --bg --set-path=/t http://127.0.0.1:8787/ttailscale funnel --bg --set-path=/m http://127.0.0.1:8787/mmousse-tree init --host <your-machine>.<tailnet>.ts.netsystemctl --user restart mousse-tree

Important: Do not expose the admin UI publicly

Never expose the /ui admin path via Funnel. The admin page (which includes memory browsing, editing, approval of Needs your approval items, and profile editing) should only be accessed from localhost, via tailscale serve (restricted to devices on your tailnet), or over an SSH tunnel. To hold writes from external platforms in Needs your approval first, set "require_confirm": true in ~/.mousse-tree/config.json.

4. View endpoint URLs and connect your platforms

Run the following command in your server terminal to print the dedicated endpoint URLs for each platform and the local admin page link:

Terminal

mousse-tree urls

Copy the URL for each platform and follow the tables in steps 2 and 3 of "Connect Your AIs to the Memory Tree" to enter the URL and custom instructions into Claude, ChatGPT, Gemini, Notion, or Claude Code.

Verification

The standalone memory tree service is running properly when all of the following are true:

Next steps

•Connect Your AIs to the Memory Tree: View the exact connector configuration paths and custom instructions for each platform.
•Memory Tree: Learn about the design of the memory tree and its five memory sources.

Last updated 2026-10-19Was this page helpful?

这一页给 agent 读的版本：在网址后加 .md 就是下面这份纯文本，和给人看的那页由同一份内容生成。Agent-readable version of this page: append .md to the URL to get the plain text below, generated from the same source as the human view.