The Mind proposes; you decide. Nothing the engine infers becomes a durable belief until you accept it, and everything you accept can be undone. Two surfaces carry that work.
Organize: the review queue
Open Organize and you get three panels: Review, Guide, and Activity.
Review holds the engine's proposals, grouped by kind, each group collapsible with a count:
- New people and projects it noticed recurring in your notes and cards.
- Merges: two names that look like the same real thing. "Rab looks like shorthand for Rabee, same person?" A merge folds the two hubs into one and keeps every mention. This includes cross-type cases, where the same thing was filed once as a person and once as a project.
- Connections between items: this note and that card are about the same work.
- Tags worth adding.
Accept a row and it becomes part of the graph. Dismiss it and it goes away. Accept all handles a whole group at once, which is the sane way through a backlog. The engine is deliberately conservative about merges: a wrong merge is worse than a missed one, so uncertain pairs stay questions.
Guide is where you teach it. Guidance rules are plain sentences ("Iris is my AI agent, not a person") the engine reads on every pass. Use it when the same wrong suggestion keeps coming back.
Activity shows what changed: what was applied, when, and the cross-links that came from it, with an undo hint for recent actions.
People: the directory the graph built
People shows every person, project and organisation the workspace knows, as a grid. Open one and its sheet shows what the Mind holds: the owner note you can edit directly, its facts, where it appears, who it is connected to and why, and the controls (merge into another hub, hide from the graph).
Editing the owner note is the fastest way to correct the record, and it teaches the assistant at the same time. Hidden entities stay hidden; merges survive re-indexing.
What to actually do, practically
Once a week is plenty. Open Organize, collapse-all, scan the counts, accept the merge questions that are obviously right, dismiss the noise, and add one guidance rule if something keeps recurring wrongly. Five minutes. The graph gets sharper and the assistant's answers get sharper with it, because chat recall reads the same graph you just cleaned.