A large codebase without drowning the agent.

Morion indexes notes, not source code. The code stays in the repo, where your agent already reads it with its own tools. Morion holds the layer above the code: architecture notes, decisions, lessons, specs, rules, and the tickets themselves. A big codebase loses that layer first, and an agent cannot rebuild it by grepping.

Create one project for the codebase. Inside it, create one folder per kind of knowledge. Open each folder’s settings. On the General tab, set Role in project: architecture, lessons, specs, rules, or the default general. The role tells Mo which folders to search when it builds a work packet. Then, on the AI Access tab, turn AI access on.

Keep the kanban folder, your work queue, separate from the reference folders.

You do not write hundreds of notes up front. Ask the agent that already knows the repo to write the first ones: one architecture note per subsystem, in its own words, with the file paths and the invariants a change must keep. Ten to twenty notes is a good first pass.

From then on, notes appear at the moments they matter. A decision becomes a note when you make it. A lesson becomes a note when a fix lands. A spec becomes a note before the ticket. The agent writes them over MCP as part of finishing work; the Morion skill tells it to.

Now a ticket is more than a title. When an agent takes it with mo_take_task, Mo returns a work packet: the ticket body, the comments that matter, the project brief, the related architecture notes, lessons, specs and rules from the reference folders, and the topics it belongs to. The agent reads three notes instead of three hundred files.

Auto-code does the same at the mo_stage gate, so the implementer in a workflow starts with the same slice a human would hand it.

On a broad ticket, a migration or a redesign, ask for more: mo_take_task takes relatedCount up to 20. If no folder has a role, the packet says so and searches the whole project instead.

Before a change that touches an old decision, ask Mo from Claude Code: it calls mo_ask, which runs on your login. The answer cites the notes and marks each claim: decided, provisional, superseded, unknown. A conflict between two notes comes back as a conflict. A gap is a gap. You get the state of the knowledge, not a confident summary of it.

In the app, the same question works in Mo Assistant, the in-app chat, on Pro with a provider key.

Rules that must load in every session stay in the file: coding principles, the verification commands, the kanban etiquette (claim, comment, move to Review). Everything historical moves out: shipped-feature recaps, postmortems, old plans. The file ends with a table that says which Morion folder holds what.

Keep it short enough to load in every session. The CLAUDE.md use case has the snippet.

Every ticket in the queue arrives with its slice of the codebase: the architecture note, the decision behind it, the lesson from last time. The agent reads three notes, not the repository. When it finishes, it writes the lesson back, so the next ticket starts one note richer.

CLAUDE.md stays short. Ask Mo tells you what is decided and what is still open, with the note it came from, and your agent gets the same answer over MCP.