Use cases

From CLAUDE.md to a memory that cites.

Your CLAUDE.md started as rules and became a diary. Split it back: rules stay in the file, everything with a date moves to Morion, and the file ends with a pointer table your agent follows. Fifteen minutes, one prompt, and the file is short again.

15 minClaude Code · Ask Mo · work packetsworks the same for AGENTS.md

1Sort the file into three kinds

Read your CLAUDE.md once and mark each paragraph. Every paragraph is one of three things: a rule, a piece of history, or reference.

Only the rules have to load every session. The rest is true and useful, and it does not belong at the top of every prompt.

Rules: stays in the file
Coding principles, verification commands, what never to touch, the etiquette for the board. Must load every session; short.
History: moves to Morion
"We shipped X on Tuesday", postmortems, old plans, why the auth module looks like that. True, useful, and not needed at the top of every prompt.
Reference: moves to Morion, gets a role
Architecture notes, decisions, lessons, specs. These become folders with a role in the project, so work packets can pull the right ones per ticket.

2Create the folders

Open the project that holds the repository. Create three folders: Architecture Notes, Lessons, Decisions. In Folder Settings → General, set Role in project: architecture for Architecture Notes, lessons for Lessons. Decisions can keep the default role, general; Ask Mo searches it either way. Then turn on AI access on the AI Access tab so the notes get indexed.

The role is what makes work packets useful later. When an agent picks up a ticket, Mo pulls the notes that fit that ticket, not the whole archive.

3Hand the split to the agent

Hand the split to the agent that wrote most of the file. Tell it what goes where, and ask to see the new CLAUDE.md before it writes anything. It creates one note per topic over MCP and tells you what it moved.

you, in Claude Code · example session Read CLAUDE.md. Move every dated or historical paragraph into Morion: architecture explanations → folder "Architecture Notes" (one note per subsystem), lessons → "Lessons" (one note per lesson, with the date), decisions → "Decisions" (one per decision: what, why, alternatives). Leave rules in the file. Show me the new CLAUDE.md before writing it. claude notes_create ×23 across three folders Proposed CLAUDE.md: 212 lines (was 1,180). Diff below.

Read the diff before you accept it. The notes already exist in Morion; the file changes only when you say so.

4End the file with a pointer table

End the file with a short section that tells the agent where history lives and how to behave on the board. Here is an example you can adapt, trimmed.

CLAUDE.md · last section ## Where to look Don't grep this file for history; search Morion. | Folder | Folder ID | What's in it | | Architecture Notes | 01M… | subsystem deep-dives, invariants | | Lessons | 01M… | one atomic lesson per note, dated | | Decisions | 01M… | what we chose, why, what we rejected | | Features (kanban) | 01M… | the work queue | ## Morion etiquette - Pick up work with mo_take_task {claim: true}; never start a todo card without claiming it. - Log progress as comments on the ticket, not in chat. - When done: tasks_move {status: "review", message: "…"}. Done is the human's column. - Before touching an old decision, mo_ask and read the citations. - After a fix that taught us something, write one note to Lessons. ## Size caps CLAUDE.md ≤ 280 lines. If you are adding more than 30 lines, the content belongs in Morion.

Put the folder id next to each name. The agent then opens the folder directly instead of searching for it.

The size caps are the part people skip and then regret. Without them the file grows back.

5Ask a question, take a task

Open Claude Code. The session starts with a short file instead of a long one. Ask what you decided about session tokens. The answer names the note each claim came from and puts a status on it. Where the notes say nothing, the answer says so.

Claude Code · ~/appMCP · morion
What did we decide about session tokens? Cite it.
mo_askproject: app
claude
  • Decision: token storage: refresh token in the OS keychain, never in the webviewdecided
  • Architecture: auth service: 15-minute access tokens, rotated on refreshdecided
  • Decision: crash reports: tokens never go to localStorage; one leaked into a crash report in Junedecided

Gap: nothing says how long a refresh token lives.

Then put it to work.

you → Claude CodeTake the next task.

The agent gets a work packet: the ticket and the two architecture notes that matter for it. The file stops being the memory. It becomes the index to one.

6Questions

What belongs in CLAUDE.md?

Rules that must load every session: coding principles, verification commands, what never to touch. History, decisions and reference move to notes the agent can search.

Does this work with AGENTS.md and Codex?

Yes. The split and the pointer table are the same; Codex reads the notes through the same MCP server.

7What you have now

A short CLAUDE.md that stays short, because the size cap tells the agent where new text goes. A memory that grows one note at a time, that any of your agents can query, and that tells you when it does not know.

next10 min
Mo on the subscription you already pay for

Run Mo on the login you already have.
next20 min
Running an epic

Put the first epic on the board.
Download