Demo workspace

A lightweight wiki for humans, a durable workspace for agents

AI Meditations separates long-term knowledge into four stable layers: raw sources, compiled wiki, operation journal, and user context. Humans get clean documents; agents get a governed MCP file workspace.

Open app Connect MCP in 3 minutes

sources/ Raw material, meeting notes, long chats, and external references.

docs/ Stable wiki pages, concepts, decisions, and compiled knowledge.

journal/ Ingests, queries, lint passes, document changes, and daily timelines.

self/ User-visible, durable preferences and working context.

Agent ingest

Preserve sources first, compile knowledge second

When an agent receives new material, it keeps the raw source in sources/ first, then promotes stable conclusions into docs/.

sources/2026/05/2026-05-01-llm-wiki.md
docs/sub_docs/ai-meditations/sub_docs/product/README.md
journal/2026/05/2026-05-01.md

Human reader

Folders are documents

docs/sub_docs/<slug>/ folders are top-level documents. Each directory is a document package: README.md is the only body, sub_docs/ contains child documents, and _attachments/ stores files owned by the current document. It stays readable for people and editable by agents.

docs/sub_docs/ai-meditations/
  README.md
  _attachments/
  sub_docs/
    product/
      README.md
    dev/
      README.md
      sub_docs/
        decisions/
          README.md

Why MCP

Switch agents without losing context

The same workspace can be exposed to different agents. Each agent starts from AGENTS.md, ai-meditations://llms.txt, and ai-meditations://skills/wiki-maintenance, then writes under the same rules.

Plain folder	AI Meditations
Files drift and agents do not know where to put things.	The top-level layers have meaning: sources, docs, journal, self.
Every agent re-learns your background from scratch.	Every agent begins with the same schema and LLM Wiki Map.
Raw notes, summaries, and logs get mixed together.	Raw sources and compiled knowledge stay separate.

中文

人的轻量 wiki，Agent 的长期工作区

AI 沉思录把长期知识拆成四个稳定层：原始资料、编译后的 wiki、操作时间线和用户上下文。人看到的是清爽文档，Agent 看到的是一套有规则的 MCP 文件工作区。

文件夹即文档

一个目录就是一个文档包。README.md 是唯一正文，sub_docs/ 里放子文档。 _attachments/ 存放当前文档的附件，不算子文档。这样人能阅读，Agent 也能用普通文件操作维护。

换 Agent，不丢上下文

同一个工作区可以暴露给不同 Agent。每个 Agent 进入时先读取 AGENTS.md、 ai-meditations://llms.txt 和 ai-meditations://skills/wiki-maintenance，再按同一套规则写入。