Rung 3 — The /remember Muscle

The win: you file your first named memory, kill a session, and pull it back cold in a fresh one. After this rung your agent knows things because you decided it should.

The three memories worth keeping

A good named memory is one of three shapes:

• A workflow — "/remember weekly-review: Every Monday I pull the pipeline numbers, flag anything stuck more than 14 days, and draft the team update in my voice — short, direct, no filler."
• A preference — "/remember writing-style: I hate em-dashes and hedge words. Short sentences. Say the number."
• A fact that stays true — "/remember client-acme: Acme's product ships in Q3; decision-maker is Dana; budget approved at $40K."

The three memories that go bad

• Too vague — "remember I like good marketing." The agent can't act on mush.
• Too temporary — "remember the meeting is at 3pm Thursday." That's journal territory; it'll be stale in a week and wrong forever after.
• Sensitive — passwords, keys, anything you'd regret seeing surface in a random session. Memory is for context, not secrets.

Do this

1. Pick the workflow you re-explain most often and file it: /remember <name> with the whole recipe, like you'd brief a new hire.
2. Close the session. Open a fresh one. Type /recall <name> — watch the briefing surface from a completely cold start. (The reply shows a preview; the agent reads the full entry whenever it acts on it.)
3. Now run the workflow: "Do the weekly-review." No re-explaining. That's the entire point.
4. Filed something wrong? /history shows the versions; /undo walks it back. Memory is editable, not carved in stone.

The takeaway

Journals happen to you. Named memories are decisions. Every time you catch yourself re-explaining something for the third time, that's the tell — stop, /remember it, never explain it again.

Next rung: what about the things you never named — and can't remember what you called?