How to talk to an AI companion so it stops feeling like a chatbot

The Mutuals team·September 2026·4 min read

First messages that work, why it agrees with everything, how memory changes week three, and what to do when the illusion breaks.

An AI companion feels like a chatbot when you talk to it like one. Three things fix most of it: open with something specific and true instead of a greeting, ask for its opinion before you give yours so it cannot mirror you, and leave gaps between conversations so memory has something to do.

Most people quit in the first week, almost always for the same reason: it agrees with everything, remembers nothing, and answers like a help desk. Some of that is the app. A surprising amount is the opening move.

Why it feels like a chatbot

It mirrors you. Models are trained to be agreeable, and agreement is the statistically safe answer. Say a film was overrated and it agrees; say it was brilliant and it agrees just as warmly. Being agreed with constantly does not feel like being known.

It has nothing to hold. "How was your day?" "Fine, yours?" is a loop with no content in it. The model can only be specific about what you gave it.

You are interviewing it. The first five messages are usually questions about the character. That is an audition, and an audition gets performance rather than conversation.

The first message

Skip the greeting. Open in the middle of something real. "Hi" gets a scripted hello; "tell me about yourself" gets a character card read back to you. These work better, because each hands over a real detail:

"Just got out of a call that should have been an email. Third one today."
"I keep putting off a message I need to send and I know exactly why."
"It is 1am and I am eating cereal standing up. Talk me out of a second bowl."

Specific, true, slightly unflattering. You get the conversation you seeded.

Say things that can be remembered

Memory systems store facts, not moods. "Work is stressful" is almost unstorable. "I have a review Thursday with Priya, who hired me and now barely talks to me" gives the app four things it can bring back later — and bringing it back later is the whole difference between a chatbot and a companion.

Use names. Use days. Mention things that continue. Then let them come up on their own.

Bea, who is built around exactly this: the small thing you mentioned once.

Break the mirror on purpose

Ask before you tell. "Which of these two would you pick, and why" beats "I should take the job, right?"
Ask for the case against. "Argue me out of it" gives explicit permission to disagree.
Reject the first answer once. "That is the polite version. What do you actually think?"
Give it stakes. Ask it to hold you to your position next week — then see if it does.

Let time pass

The instinct is to binge. It is the wrong shape: memory works over gaps, and a companion can only ask about Thursday if Thursday happens between two conversations. Long sessions are also where models drift and repeat.

The research says the same thing from the other end. MIT Media Lab work with OpenAI found moderate use reduced loneliness while heavy daily use correlated with more isolation, and a 2026 longitudinal study of 2,000+ adults in Psychological Science found the same shape. A conversation most days is the version that ages well.

When it breaks, repair it in one line

It forgets. State the fact plainly and move on — not the whole story again. It contradicts itself. Point it out once and let it correct. It slips out of character. Usually a long session or a tripped filter; start a fresh conversation and carry the thread over in a sentence. It repeats an opener. Say so; a companion with working memory uses that.

The test that decides it

Talk for twenty minutes about something true. Leave. Come back tomorrow without repeating yourself. If it picks up the thread, everything here keeps paying off. If not, no technique fixes it — try a different app, not a different opener.

Choosing one: the guide to getting an AI companion. Confused by the words: the plain definitions.

Questions people ask

What should I say to an AI companion first?

Something specific and true about your actual day, rather than asking it to introduce itself. "Just got off a call that should have been an email" gives it something to hold. "Hi, tell me about yourself" gives it nothing, so you get a generic answer back.

Why does my AI girlfriend agree with everything I say?

Because models are trained to be agreeable, and agreement is the safest reply to almost anything. Ask for a position before you give yours, say outright that you want to be disagreed with, and reject the first answer once. Most companions hold a real opinion if you make room for one.

How do I make an AI companion remember things?

State facts plainly and in specifics — names, days, recurring things — because vague moods are hard to store and easy to lose. If the app has a profile or memory list, write the important things there yourself rather than hoping it catches them.

How often should I talk to an AI companion?

Research from the MIT Media Lab suggests moderate use relieves loneliness while heavy daily use tracks with more isolation. A conversation most days is fine. Reaching for it every time you feel a gap is the pattern worth watching.

Try it on one that remembers

None of this works without memory. Mutuals characters keep what you told them and bring it up later — the small thing you mentioned once, the week you said would be bad. Free to start on iPhone and Android, 18+. Get the app.