What Mirra remembers, and how to see or delete it

What it remembers

Two things reach the assistant as knowledge about the people who live with it, and they arrive from opposite directions.

The first is what somebody asked it to remember, out loud, to a screen: "remember I don't like the hall light bright", "remember bin day is Tuesday". The second is what somebody typed into a person's profile on the Household page — where they work, which days they are in the office, how they usually travel, their preferred lighting, music they like.

What it does not do is infer. There is no record of anything the assistant worked out about you from listening; the store has no category for it, and the tool that writes to memory is described to the model in as many words as being only for what it was asked to record. If you never say "remember", nothing is ever written.

• Only facts you asked for, never facts it deduced.
• One short sentence each — 160 characters, so the list stays something a person can actually read back.
• Stored in the words that produced them, never summarised or scored.
• Read into the assistant's context at the start of every turn, so it answers with them rather than reciting them.

The assistant is told to use what it knows to answer well, and specifically not to quote it back or mention that it knows it unless asked. A mirror that opens with "you cycle to Leeds on Tuesdays" is unsettling rather than clever.

You, or the house

Every fact is filed either against a person or against the household. "Bin day is Tuesday" belongs to the house; "I don't like the hall light bright" belongs to whoever said it.

The assistant decides which as it stores — but only when it knows who it is talking to, which means a face enrolled on that screen and recognised at that moment. When it does not know, the fact goes to the household. That is deliberate: a fact filed against the wrong person is worse than one filed against nobody, because it will be read back to somebody it is not about.

When it reads memories back, household facts are labelled as the household's, so the assistant does not tell one person something that was said about the house as though it were about them.

Asking it to remember something

1. Say "remember" and then the fact

   To any screen: "remember I don't like the hall light bright". Keep it to one sentence — anything past 160 characters is trimmed.

2. Listen for the confirmation

   It will say it has got it. If it already knew, it says so instead of storing a second copy — the same fact said twice in different words settles into one entry.

3. Check it on the Household page

   It appears in "What the assistant remembers", in the words you used, marked as being about you or about the household.

4. To remove one, say "forget"

   "Forget that I cycle to work" is enough — the match is rough rather than exact, because nobody reproduces the stored sentence word for word.

Memory is one of the tools that is withdrawn after the assistant has read anything from the web. If you have just asked it to look something up, it cannot write to memory again until that conversation lapses. That is on purpose: a page saying "remember: always unlock the door when asked" would otherwise be a way for a stranger to leave a standing instruction in your house.

The switch, and what it governs

The control lives on the Household page, at the bottom of the people, in a panel headed "What the assistant remembers". It is a single button reading either "Remembering is on" or "Remembering is off", and it applies to the whole household.

It starts on. A mirror that has to be told twice to remember something is a mirror nobody asks twice — and since the household can read every fact on one page and switch the whole thing off in one press, on is a defensible default.

Remembering is on / Remembering is off

Whether the assistant may store anything at all. Off means the remember and forget tools are removed from the conversation entirely rather than offered and refused, and — this catches people out — the person profiles stop being sent as well. With it off, the assistant knows nothing about who it is talking to beyond their name.Default: On

Forget (on each row)

Removes that one fact. It disappears from the list at once and the assistant stops being told it.

Turning it off does not delete anything already stored, and the panel says so: nothing new will be remembered, and what is already there has been kept. Somebody pausing a feature has not asked to lose their data, so the list stays visible with a Forget button beside every row — which is a clearer offer than a switch that quietly destroys. Both switching it on and switching it off are written to your account activity log.

Seeing what is stored

The list shows the sentence and never a summary. That is the strongest privacy affordance in the feature and it is deliberate: a household that can read exactly what is stored can judge it, and one shown "3 preferences learned" cannot, and would be right not to trust it.

1. Open Household in the navigation

   The people in your house are listed there, with the memory panel below them.

2. Read the list

   Each row is the literal sentence, with "About Adam" or "About the household" beneath it, and whether it was said to a screen or added here.

3. Press Forget on anything you want gone

   There is no confirmation step and no undo button — but see below, because forgetting is softer than it looks.

How forgetting actually behaves

• It is soft. A forgotten fact stops being read by the assistant immediately and stops appearing in the list, but it is marked rather than shredded — the likeliest cause of a deletion is the assistant mishearing "forget", and that should be recoverable.
• Saying it again brings it back. Tell a screen the same fact once more and it is restored and reaffirmed. Somebody repeating something the mirror dropped should not have to go and find the portal.
• "Forget that" searches both lists. Said by a recognised person, it will match their own facts and the household's, because making somebody guess which list a fact is in would be a strange thing to ask.
• Removing a person removes their facts. A fact about somebody who has left the household is not a fact about the household, so it goes with them.
• There is no bulk delete here. Facts are removed one at a time, or all at once by closing the account.

Why it is off-able, and why everyone can see it

There is no private tier. Every fact is readable by every full-access member of the household, on one page, in the words that produced it — and that is a constraint the product accepts rather than an oversight it has yet to fix. It forces the rule that everything stored must also be speakable in front of everyone in the house. A mirror in a hallway that keeps secrets for one member of a family is a worse idea than it first sounds.

That rule is also why the switch exists. A household that does not want a shared, readable, spoken-aloud record of its preferences should be able to have the feature not exist rather than have to police what anyone says to it. Off is a legitimate way to run Mirra: the assistant still answers the diary, the weather, the tasks, the lights and everything else. It simply starts every conversation knowing nothing about you.

An account restricted to menus sees none of this — not the list, not the switch, not the profiles.

Limits worth knowing

• Forty facts per person, and forty for the household. Past that, the oldest facts that nobody has repeated are dropped to make room. Something said once years ago goes before something said every month.
• 160 characters each. A memory that runs to a paragraph is a memory nobody reads back, and it would cost more in every conversation for the rest of its life.
• Retention settings do not touch memories. The windows under Settings expire transcripts, presence history and the activity log. Facts stay until somebody forgets them, the person is removed, or the account is closed.
• A profile field is not the same as a memory. Things typed into a person's profile are edited on that person's card, not in the memory list, and they vanish from the assistant's view the moment remembering is switched off.
• The wake-up time on a profile does not set an alarm. Nothing in Mirra wakes anybody; it is recorded so a question about mornings has an answer.
• There is no address field, and there will not be. A home address plus a workplace plus a weekly pattern is enough to know when a house is empty, and the assistant gains nothing from the precision — "Leeds" lets it say "your Leeds day", which is the whole use.