What should an AI assistant be allowed to remember about you?
An assistant should remember what you asked it to keep, say so when it stores a new fact, and drop that fact on demand. It should not infer health, religion, politics, sexuality or ethnicity. GROX notes a new fact under the answer, removes it if you type forget that, and will not guess those traits. Silent learning erodes trust because you cannot audit what is stored.
Why does silent learning erode trust?
Memory that is never announced is still memory. The next reply may use a job title, a city or a preference you never agreed to keep. You cannot correct what you cannot see. Over time the gap between what you think the system knows and what it actually uses becomes the real product, not the chat.
A simpler notes app is often the better choice when you only need a list you wrote yourself. An assistant that acts across email, calendar and chat is different: it will pick up facts unless it is forced to declare them. The trade-off is convenience against a trail you can read and undo.
What may it infer, and what must it keep only when you say it?
Routine work facts — a nickname you used, a project name, a preferred tone — can be inferred from what you type, if the system tells you it did so. Traits that are easy to stereotype must not be filled in from context. Health, religion, politics, sexuality and ethnicity belong in memory only when you state them about yourself.
People notes drawn from mail and meetings should stay short: who someone is and what was last agreed, never a guess and never a sensitive trait. If a brief needs none of that, the note should have its own off switch.
- Spoken learning
- The system states, under the answer, that it kept a fact, so you can undo it in the same turn.
- Sensitive traits
- Health, religion, politics, sexuality and ethnicity are stored only if you say them about yourself.
- App-scoped forget
- A fact learned from one connected app can be dropped without wiping the rest of memory.
- People notes
- One short note per person from mail and meetings; never a guess; own switch in Settings.
How does GROX handle what it remembers?
When GROX learns something about you it says so under the answer. Typing forget that removes it. It never guesses health, religion, politics, sexuality or ethnicity; those stay only when you say them about yourself. What it learned from one app can be forgotten on its own, in the chat or in Settings → Memory.
People notes come from emails and meetings used to prepare meeting briefs: who they are and what was last agreed. They have an off switch in Settings → Memory, and the next brief uses them. Nothing in that path invents a sensitive trait.
- Look under the answer for the line that a fact was kept.
- Type forget that to drop that fact at once.
- Use Settings → Memory, or the chat, to drop what one app taught.
- Turn people notes off if briefs should not reuse them.
How do memory styles compare in practice?
The useful split is not “smart versus dumb”. It is whether you can see the store, undo a single fact, and keep protected categories empty unless you filled them. Tools that only keep a local scratchpad avoid the problem by remembering almost nothing; tools that act across inboxes need the spoken-learning pattern.
| Approach | What is stored | How you undo it |
|---|---|---|
| Local notes only | What you typed into the note | You edit or delete the note |
| Silent profile | Inferred facts with no notice | Buried settings, if any |
| GROX | Facts it states under the answer; no guessed sensitive traits; people notes without guesses | forget that; forget one app in chat or Settings → Memory |
Common questions
Should an AI tell me when it learns something?
Yes. A line under the answer is enough. If the system never says it stored a fact, you cannot audit or correct it. Spoken learning is the minimum for anything that later changes a reply, a brief or a routine.
What should never be inferred about me?
Health, religion, politics, sexuality and ethnicity. Those should be kept only when you state them about yourself. A meeting brief or people note should not fill them in from tone, names or mail headers.
How does GROX let me undo memory?
GROX puts a note under the answer when it learns something. Type forget that to remove it. Facts from one app can be forgotten on their own in the chat or in Settings → Memory. People notes have a separate off switch there.
Is a simpler tool better than an assistant with memory?
If you only need a list you wrote, a notes app is cleaner. An assistant that reads mail and calendars will pick up facts unless it declares them and lets you drop them. Choose spoken learning, not a hidden profile.
Read how memory and people notes work in the Help Centre or start from Home.