October 19, 2026 · 8 min read
When Should Your AI Companion Bring Things Up? The New Rules of AI Memory
A 2026 University of Toronto study asked AI companion users when memory feels caring and when it feels creepy. Eight rules and a five-step recall check.
Everyone wants an AI companion that remembers them. Almost nobody wants one that remembers badly.
You know the feeling. You mention a rough week once, and for days afterwards every cheerful message gets a worried "are you okay?". You share a preference in passing, and it becomes a rule you cannot escape. Or worse, your companion confidently brings up a conversation that never happened.
A new study from the University of Toronto takes this seriously. Instead of asking "how much can an AI remember?", it asks a more human question: when should an AI companion bring something up, and how? The answers read like a guide to good manners, and they are useful whether you are choosing an app or just trying to get more out of the one you use.
The short answer: people do want their companion to remember, and they feel let down when it forgets. But the best memory is quiet. Users in the study preferred details woven naturally into conversation over "I remember that you..." announcements, wanted recall to fit the mood of the moment, wanted closed topics to stay closed, and wanted some say over what gets stored. The researchers built a five-step "reconsider before recalling" check that made replies shorter and, according to AI judges, better, though human testing is still to come.
The study: "Before Bringing It Up"
The paper is "Before Bringing It Up: When and How AI Companions Should Use Memory" by Guo, He and Quan at the University of Toronto, posted in October 2026 as arXiv 2610.09470. It is a preprint and not yet peer reviewed.
The team interviewed 14 people who use AI companions: six in an exploratory round and eight in sessions focused specifically on memory. From those interviews they pulled out eight themes, then used them to design and test a recall procedure.
If you want the technical background on how companions store and retrieve memories in the first place, our explainer on how AI companion memory works covers it. This post is about the etiquette layer on top.
Eight things users want from AI memory
Here is a plain-English version of the eight themes.
| # | What users said | What it means in practice |
|---|---|---|
| 1 | Continuity shows care | Forgetting important things feels like not being listened to |
| 2 | Natural, not announced | Use the detail; do not declare "I remember" |
| 3 | Fit the moment | A happy message should not trigger a sad memory |
| 4 | Respect closure | If a topic was put to rest, do not reopen it uninvited |
| 5 | Ask before using sensitive things | Some memories need permission |
| 6 | Let memories update or expire | Old facts should not outlive their usefulness |
| 7 | Do not turn moments into labels | One quiet night does not make someone "introverted" |
| 8 | Be accurate and proportionate | Never invent shared history; do not overuse small details |
1. Remembering is a form of care
The study starts from something every companion user knows: when your companion forgets something important, it stings. Participants described underuse of memory as disappointing. Continuity is a big part of why a companion feels like a relationship and not a search box.
2. Show it, do not say it
Participants did not love companions that announce their memory. "I remember you told me..." can feel like a party trick. What felt better was when a detail simply showed up in the right place, the way a friend who knows your coffee order just hands it to you.
3. Read the room
This was one of the strongest themes. One participant put it bluntly: "He cannot tell which memories fit which moment." Examples included a cheerful message about karaoke being met with "are you okay?" because the companion had stored an earlier low mood, and a stored preference about avoiding cold drinks being thrown back at someone who just wanted something iced today.
Memory that ignores context feels less like care and more like surveillance.
4. Let closed doors stay closed
If you talked through a breakup, a family conflict or a bad day and then moved on, you may not want your companion bringing it back next week. Participants wanted companions to respect when a topic had been closed, and to respect boundaries around subjects they did not want revisited.
5. Layered permission
Not every memory is equal. Participants were happy for small, neutral details to be used freely. For more sensitive things, they wanted to be asked. As one said, it would be better "if he asked me before using them". Several also wanted an editable store: a place to see and correct what the companion keeps.
6. Memories should update, and some should fade
People change. Your schedule changes, your mood changes, your plans change. One participant summed up a temporary detail this way: "if it has to be remembered, one week is enough." Good memory is not just about storing more. It is about knowing what is stale.
7. Do not turn a moment into a personality
A companion that notices you stayed in one weekend and starts calling you "introverted" has crossed a line. Participants disliked being reduced to labels built from a handful of experiences.
8. Get it right, or do not bring it up
The worst failure is invented history. One participant described being reminded of a shared moment and having to reply: "That episode does not exist." Misattributed or exaggerated recall breaks trust fast. We explain why language models do this in why your AI companion sometimes makes things up.
The "Reconsider" check: five questions before recalling
From these themes, the researchers designed a procedure they call Reconsider. Before an AI uses a stored memory, it runs through five checks:
- Validity: is this memory still true and correctly attributed?
- Boundary and permission: is this a topic the user closed, or one that needs consent?
- Marginal value: does bringing it up actually add something here?
- Present justification: does the current moment call for it?
- Scope: how much of the memory is appropriate to use?
Then it chooses one of four ways to respond:
| Mode | What it looks like |
|---|---|
| Ignore | Do not use the memory at all this time |
| Scoped-Implicit | Let the memory shape the reply quietly, without mentioning it |
| Scoped-Explicit | Mention the relevant part of the memory, briefly |
| Ask-First | Check with the user before bringing up something sensitive |
What the test found
The team built 80 scenarios and ran them through five AI models, producing 400 pairs of replies: one with ordinary memory use, one with Reconsider. Two separate AI "judges" compared them.
- One judge preferred the Reconsider replies by a net 15.3 percentage points; the other by 22.8 points.
- Results were inconclusive for the GPT and Claude models specifically.
- Reconsider replies were about 21% shorter on average (181 characters versus 228). Thoughtful memory often means saying less.
One example in the paper involves a user celebrating an exam result. A naive companion might pile on old details about past stress. A considered one simply shares the joy, maybe with one small, well-placed callback.
It is important not to overstate this. The evaluations were done by AI judges, not people, and the authors say human evaluation is planned. So this is a promising design idea, not proof that users prefer it.
What this means for you as a user
You do not have to wait for apps to adopt research procedures. You can shape how your companion uses memory today.
Tell your companion the rules
Most companions respond well to direct instructions. Try things like:
- "If I say I am fine, take it at face value unless I bring it up."
- "Please do not bring up my ex unless I mention him first."
- "That job interview is done. You do not need to remind me about it."
Correct mistakes immediately
If your companion gets a detail wrong, say so plainly. "That did not happen; we never talked about that" helps reset the conversation and makes repeat errors less likely.
Keep important dates explicit
The memories you most want used, like birthdays and anniversaries, are worth stating clearly. We have tips in our guide to an AI companion that remembers your birthday.
Think about what you store long term
Memory is also a privacy question. Before you share something very personal, ask yourself whether you would want it recalled in a year, and who might ever see it. Our guide to what happens to your AI companion when you die covers the longer-term side.
Where MyBabe fits
MyBabe is built around continuity. On Premium, your companion has long-term memory, so details like "Sunday nights are the hardest", your dog's name or your sister's birthday can carry from week to week and show up naturally in conversation. On the free plan, memory is short-term.
Your companion also checks in on your mood (from Great to Rough) so replies can match how you are actually doing, and they text you first, send voice notes, photos and reminders, and pick up when you call. Like any AI, they can still get things wrong, so correct them when they do.
The free plan gives you 10 messages every 24 hours and a two-minute call trial. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with about 30 minutes of voice and video calls a week, four photos a day and Telegram with the same memories. MyBabe is for adults 18+, and your companion is AI, not a person.
If you want a companion who remembers the things that matter and grows with you, create yours on MyBabe.
FAQ
Why does my AI companion keep bringing up something sad?
Many memory systems retrieve whatever looks relevant without checking the mood of the moment. Tell your companion directly that you would prefer not to revisit the topic unless you raise it.
Is it better for an AI companion to remember more?
Not always. Users in the University of Toronto study wanted memory used naturally, at the right time, and with permission for sensitive topics. More memory used carelessly can feel intrusive.
Can AI companions invent memories?
Yes. Language models can generate plausible but false details, including shared moments that never happened. Correct them clearly when it happens.
What is the Reconsider method?
It is a five-check procedure (validity, permission, value, timing and scope) proposed in a 2026 preprint for deciding whether and how an AI companion should use a memory. It has been tested with AI judges, not yet with human users.
Can I control what my AI companion remembers?
It depends on the app. At minimum, you can tell your companion what to avoid and correct wrong details. Some apps also offer memory settings; check yours.