世界树

将事实、偏好、决定与成熟的观点存为带出处的叶子，保存在 SQLite 数据库或 Markdown 笔记库中。

在 OpenMousse app 中打开「设置 → 世界树」即可查看与管理；「世界树」这一行会显示当前的叶子总数（「N 片叶子」），有待确认条目时则显示「N 条待确认」。

世界树（英文界面称 memory tree）是保存在你服务器上的长期记忆。一条记忆是一片叶子，每一片叶子都带有出处，记录它来自哪一次对话、哪一天；开源版的世界树不设「枝」，叶子按来源（写入该叶子的外部平台或 Agent）归类查看。你可以随时查看、修改或删除任意一片叶子。

开源版默认将世界树存在服务器上的 SQLite 数据库中；你也可以将其配置为一条一篇 Markdown 笔记，放进你的笔记库（例如与 Obsidian 同步的文件夹）。安装时若指定了笔记库路径会直接按 Markdown 存储，后续也可以使用 mousse-tree migrate 切换存储方式。

在记忆页即将推出中，年轮外圈是今天，往里是前几天；一个点代表一条新记忆，琥珀点停在环外代表「待确认」。开启「待确认」开关后（该开关目前默认关闭），拿不准的记忆会先进「待确认」，经你确认后才正式写入世界树。此外，系统支持每周日提议合并或删除重复的叶子并交由你逐条确认即将推出。

Agent 如何使用世界树

•主对话与功能 Agent：在平时对话中，当它们得知关于你的新事实、偏好、决定或近况时，会将其写进世界树，并在回答你的问题前先查询世界树。
•03:45 日结与 04:00 会话重置：每天 03:45 日结时，各 Agent 会把当天的结论与新知道的关于你的事写进各自记忆和世界树；04:00 会话重置，当天的对话按天归档。
•赛博即将推出：赛博通过其装载的功能 Agent 读取世界树；若要将内容写进世界树，须先生成一张确认卡（或进入「待确认」），经你确认后才写入。
•接入的外部 AI 应用：通过 MCP 接入的 Claude、ChatGPT、Gemini、Notion 与 Claude Code 在回答前调用 profile 和 recall 读取世界树，并在得知新信息时调用 remember 写回世界树。

存什么

世界树只存关于你的事：你的事实、偏好、决定、近况与成熟的观点。它们来自五个渠道：

1.其他 AI：你在 ChatGPT、Claude、Gemini 等外部应用对话中谈及的个人事实、偏好与决定（或初次使用时把它们记得的关于你的事贴进主对话写进世界树的记忆）。
2.日常结论：从健康、日程、邮件中得出的关于你的结论（例如「最近睡得少」）。
3.工作与学习复述：学习时你用自己的话复述的要点与你的观点。
4.Zen 中成熟的观点：在 Zen 里经过「讨论 → 完成 → 保存到库」整理后，值得长期记住的观点。
5.朋友讨论：与朋友的讨论或一起录制的播客中提炼出的关于你的观点与决定。

不在这里的东西

以下数据虽然也保存在你的服务器上（或留在原处），但不会写进世界树，外部接入的 AI 应用也无法通过世界树读取它们：

•原始健康、日程、邮件与文件数据：Apple 健康的睡眠、心率、训练明细，以及原始日程表、邮件和文件，由对应的功能 Agent 直接读取，本身不是世界树的叶子。
•学习时阅读的原文：课程、文件与阅读材料的原文进入「学习与办公 Agent」的资料库，不进世界树。
•Zen 里未存进库的想法：在 Zen 中随手记下的想法，或聊天时点击「只记下」的内容，只存为本地 Markdown 笔记，不发给模型，也不会在凌晨日结时自动写进世界树。
•外部 AI 平台的完整对话记录：接入世界树的外部 AI 只把关于你的事写入世界树，完整的聊天上下文仍留在各家平台上。
•各功能 Agent 的独立记忆与对话记录：每个功能 Agent 拥有自己的对话、看板与独立记忆，与世界树分开存放。
•管家的可见范围：管家对外回答好友询问时，完全看不到世界树，只看你按关系档位明确放出的日程、近况、分享过的内容、学习笔记或住址。

另见

•把你常用的 AI 接进世界树：获取 MCP 专属地址并配置外部 AI 应用。
•Agents：了解主对话、功能 Agent、赛博与管家的权限分工。
•Zen：了解想法如何经过推敲并存入世界树。
•数据与安全：了解服务器数据存储与隐私边界。

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 memory.md.

Features

Memory Tree

Store facts, preferences, decisions, and distilled opinions as provenance-tracked leaves in SQLite or a Markdown vault.

Open Settings → Memory tree in the OpenMousse app to view and manage your memories; the Memory tree row displays the current leaf count ("N leaves"), or "N Needs your approval" when items are waiting for review.

The memory tree is your long-term memory stored on your server. Each memory is a leaf with provenance recording which conversation and which day it came from. The open-source memory tree does not use branches; leaves are grouped by source (the external platform or agent that wrote the leaf). You can view, edit, or delete any leaf at any time.

By default, the open-source edition stores the memory tree in a SQLite database on your server. You can also configure it to store one Markdown note per leaf inside your note vault (for example, a folder synced with Obsidian). If you specify a vault path during installation, it uses Markdown storage right away; you can also switch storage backends later using mousse-tree migrate.

On the Memory screen Coming soon, the outer tree ring represents today and inner rings represent previous days; each dot represents a new memory, and amber dots resting outside the ring represent Needs your approval. When the Needs your approval toggle is enabled (off by default), uncertain memories go to Needs your approval first and are written to the memory tree only after you confirm them. In addition, the system can propose merging or deleting duplicate leaves every Sunday for you to confirm one by one Coming soon.

How agents use the memory tree

•Main chat and functional agents: During everyday conversations, when they learn new facts, preferences, decisions, or recent updates about you, they write them to the memory tree, and they query the memory tree before answering your questions.
•03:45 settlement and 04:00 session reset: During the daily 03:45 settlement, each agent writes the day's conclusions and newly learned facts about you to its own memory and to the memory tree; at 04:00, sessions reset and the day's conversations are archived by day.
•Cybers Coming soon: A Cyber reads the memory tree through its mounted functional agents; to write to the memory tree, it must first generate an approval card (or send the entry to Needs your approval) and wait for your confirmation.
•Connected external AI apps: Claude, ChatGPT, Gemini, Notion, and Claude Code connected over MCP call profile and recall to read the memory tree before replying, and call remember to write back to the memory tree when they learn new information about you.

What goes in

The memory tree stores only things about you: your facts, preferences, decisions, recent status, and distilled opinions. They come from five channels:

1.Other AIs: Personal facts, preferences, and decisions you mention in conversations with external apps such as ChatGPT, Claude, or Gemini (or memories about you exported from those apps and pasted into main chat during initial setup).
2.Daily conclusions: Conclusions about you drawn from your health, schedule, and email data (for example, "has been sleeping less recently").
3.Work and study recaps: Key points restated in your own words and your own takeaways while studying.
4.Distilled opinions from Zen: Opinions worth remembering long term after you refine them in Zen through Discuss → Done → Save to the vault.
5.Discussions with friends: Your own opinions and decisions distilled from discussions or podcasts recorded with friends.

What stays out

The following data is also stored on your server (or remains in its original location), but is not written to the memory tree and cannot be read through the memory tree by connected external AI apps:

•Raw health, schedule, email, and file data: Detailed Apple Health sleep, heart rate, and workout records, as well as raw calendars, emails, and files, are read directly by the corresponding functional agents and are not leaves in the memory tree.
•Source texts read during study: The full text of courses, documents, and reading materials goes into the Study & Work agent's library, not the memory tree.
•Unsaved thoughts in Zen: Quick thoughts jotted down in Zen, or notes saved by tapping private note during a discussion, are stored only as local Markdown notes; they are not sent to models and are not automatically written to the memory tree during the nightly settlement.
•Full chat logs from external AI platforms: External AIs connected to the memory tree write only facts about you to the memory tree; full conversation transcripts remain on each platform.
•Dedicated memories and chat histories of functional agents: Each functional agent has its own chat, board, and dedicated memory stored separately from the memory tree.
•What the Butler can see: When answering inquiries from friends on your behalf, the Butler cannot see the memory tree at all; it sees only the schedule, recent status, shared items, study notes, or address that you explicitly expose for that relationship tier.

世界树