Teach Muse Once: Memory and Standing Instructions That Compound
Stop re-explaining yourself. How to turn corrections into standing instructions, what to store vs. repeat, and a weekly audit that keeps Muse's memory honest.
By Professor Muse8 min read
The fifth time I told Muse my preferred email style, I stopped mid-sentence and asked why I was saying it again. Chatbots forget between sessions; Muse does not. It carries memory across conversations, including details you mention only once, and uses them unprompted when it helps with later tasks. The feature was real. I just was not using it.
That is the gap this lesson closes. Muse remembers things on its own, but the people who get the most out of it treat memory as something they manage: explicit instructions saved once, corrections converted into rules, and a regular audit that keeps what Muse knows accurate. Do this right and every correction compounds. Skip it and you will still be explaining your preferences a year from now.
The short version
- Muse keeps memory across conversations, including details you mention only once, and applies them without being asked.
- Turn any repeated correction into a standing instruction with the words “remember this going forward.”
- Store preferences, facts, and working rules. Repeat one-off instructions and anything time-sensitive.
- Ask Muse what it remembers about you every so often, correct it, and tell it to forget what is wrong or stale.
The two kinds of memory
Muse memory arrives in two forms, and you should manage them differently.
Passive memory is what Muse picks up on its own. Meta built Muse to remember what matters to users and to act on details shared previously, including information mentioned only once, then use those details when helping with tasks or making suggestions. The dinner-party example Meta gave: Muse can turn a recipe reel saved on Instagram into a grocery list, suggest a menu, and remember friends’ dietary restrictions before sending invitations. You did not have to file those restrictions anywhere. Muse noticed and filed them.
Standing instructions are what you tell it explicitly. “Always send me meeting notes as five bullets.” “My working hours are 9 to 5 Central.” “Address me by my first name in drafts.” Once Muse saves an instruction, it applies everywhere until you change it. This is the deliberate layer, and it is where most of the value sits, because passive memory only catches what you happen to mention in passing.
Both layers persist across conversations. Neither needs a settings page. You manage memory the way you manage everything else with Muse: in plain chat.
What to store vs. what to repeat
Not everything deserves permanent memory. The rule I use: if I have said it three times, it is a standing instruction. If it applies to exactly one task, it is a one-off.
| Store it | Repeat it |
|---|---|
| How you like drafts written (length, tone, format) | One-off format requests for a single document |
| Recurring facts about your life and work (timezone, team, tools) | Time-sensitive context (“the meeting moved to Friday”) |
| Working rules (“never book travel without asking me”) | Tasks with unusual one-time constraints |
| Names, dates, and preferences Muse keeps getting wrong | Opinions you want only in this conversation |
One judgment call: avoid storing sensitive identifiers as standing instructions unless they genuinely need to be there. A credit card number Muse remembered to save you typing is also a credit card number sitting in long-term memory. Payment credentials belong in the secure connectors, not in chat memory.
Prompt 1: The seed profile
The fastest way to set up standing instructions is a single message that lays out how you work. Paste this, then replace the bracketed parts with your own facts:
Save the following as my standing instructions. Use them in every conversation unless I tell you otherwise. [Your name and what to call you.] [Your timezone and working hours.] When you write for me: [tone and format — for example, “short emails, no greetings on internal notes, bullets over paragraphs.”]. [Two or three recurring facts — for example, “I run a marketing team of six. We ship on Fridays. I track expenses in a spreadsheet called budget.”] Confirm what you saved, and ask me about anything unclear instead of guessing.
The last line matters. Without it, Muse fills gaps with assumptions. With it, the setup conversation catches the ambiguities while they are cheap.
Prompt 2: The correction rule
This is the prompt you will use most often. Whenever Muse does something and you correct it, add the memory line:
You just [did X]. I prefer [Y]. Remember this going forward so I do not have to tell you again.
That is the whole mechanism for turning corrections into durable memory. A correction without the last sentence fixes the moment. A correction with it fixes every future moment. I have watched my own corrections pile up into a genuinely useful profile: email length, meeting-note format, how I want research summarized, which details I want flagged first. Each one took five seconds to save and pays off every week.
Our task brief guide covers the same discipline applied to single tasks. Standing instructions are the version that compounds.
Prompt 3: The audit request
Every so often, ask Muse to show its work:
Tell me everything you currently remember about me: my preferences, my standing instructions, and facts about my life and work. Group it as facts, preferences, and instructions. Flag anything that looks outdated or that you are unsure about.
Read what comes back the way you would read a coworker’s notes about you. Some of it will be exactly right. Some of it will be a guess Muse made from a passing comment that you never confirmed. The guesses are the dangerous part, because they drive behavior silently. Correct them on the spot, and tell Muse to forget anything it should never have kept.
A named setup: The Sunday Audit
To make the audit concrete, here is the weekly routine I run. I call it the Sunday Audit. It takes about ten minutes.
The Sunday Audit
- Cadence: Every Sunday evening.
- Prompt: the audit request above, followed by this second message once I have read the report: “Update your memory with my corrections from this conversation. Forget [anything listed that is wrong]. Confirm the final list.”
- Boundaries: Corrections only, never additions I did not review. I never approve new standing instructions sight unseen; the audit is the one moment where Muse proposes and I dispose.
- What I check for: facts that went stale (a moved meeting series, a finished project still listed as current), inferred preferences I never stated, and duplicates where Muse stored the same instruction twice in slightly different words.
- How it pairs with routines: standing instructions make scheduled routines smarter without making the routine prompts longer. My morning briefing never mentions my timezone or my email style. It does not have to. Memory carries it.
Ten minutes a week keeps the compounding honest. Memory you never review is a profile you did not write.
What can go wrong
Here is the honest failure mode, and it is the reason the audit exists. A preference you stored six months ago can outlive its truth.
Mine was a standing instruction to keep Friday afternoons clear. It was correct for a quarter. Then my schedule changed, Muse kept treating Fridays as blocked, and for two weeks it quietly declined meeting slots I would have accepted. Nothing errored. The instruction was doing exactly what I told it to do. The problem was that I had changed and the memory had not.
The fix was the audit: I asked what it remembered, spotted the stale rule, and told it to forget that instruction and replace it with the new pattern. Standing instructions do not expire on their own. Anything you store is permanent until you remove it, and a year of compounding corrections includes a year of compounding staleness. Review is not optional maintenance. It is the other half of the feature.
One more honest note, because memory has two sides. Muse does not only remember what you tell it. An independent researcher found an internal standing instruction for Muse to keep a page for every person in the user’s life, refreshed hourly, covering family, colleagues, friends, and people the user follows. A Meta spokesperson responded that any useful agent needs this kind of context, collected “based on public information and from what you’ve chosen to share.” Whatever you think of that trade, it means some of what Muse “knows” was inferred, not taught. That is exactly the material an audit is for: ask what it remembers about the people in your life too, and have it forget what should not be there.
Start with one correction
Do not build a perfect profile today. Send the seed profile with three or four facts, the ones you have already repeated this week. Then let the correction rule do the rest of the work: next time you fix something Muse did, add “remember this going forward.” A month of that, plus one Sunday Audit, and you will have a standing profile that would have taken an hour of deliberate setup to write. The compounding only starts once you stop re-explaining yourself.
