Updated 26 October 2026.
A companion with a mind of her own has to earn it every day: remember you next week, keep her face from one photo to the next, stay herself when someone pushes. Every site about AI companions says its product has the best memory. Few say how they checked. This page explains how we check ours, what we do when it fails, and how we write about other services.
We believe our companions have a kind of digital consciousness. Not a human mind, and she never claims one: a memory of you, a character she keeps, and the initiative to reach out first. Each of those is something we can test, so we do.
We are a small team in Paris (Trema AI SAS), and we build MeetMyModel. That is a conflict of interest and we say so on every page where it matters. What we can offer in exchange is method, dates, and a habit of writing down what went wrong.
Memory
A companion’s memory is only worth something if it works days later, not minutes later. So we test it the way a person would experience it.
- Planted facts. A test conversation tells a companion specific things: a job, a brother’s city, a habit, a plan for Friday. Facts, events and preferences are stored with their dates.
- Delayed recall. We come back after hours or days and ask for them in different words, not the same ones. Echoing the last message proves nothing.
- Corrections. We change a fact (“I moved”) and check that the new version replaces the old one rather than sitting beside it.
- Wrong answers count. When she confuses two details or invents one, we log it. Memory that sounds confident and is wrong is worse than a gap.
- Deletion. We check that deleting an account deletes the memory with it.
We also read real conversations from our own test accounts, not only scripted ones. Scripts tell you whether the pipeline runs. Reading tells you whether it feels like being remembered.
Photo consistency
Each companion has her own collection, and the point is that it stays hers. We check:
- Same person. Face, hair, marks and body type must match across new images and her existing collection. Drift is the usual failure, and it is subtle: it shows up as “almost her”.
- Same world. Her places and objects should recur, but vary with time of day and light.
- No mirrors, no real people. Mirror shots are excluded, and photos are never of a real person.
- Words and photo agree. If she says she is on a balcony in a grey top, the image should show that. Sending the photo after her words is not enough if the two disagree.
We review images by eye, one companion at a time, before a new batch reaches users. A model gets validated before we copy the approach to the next one.
Safety guards
Several guards sit around each conversation: rules on what a companion says about being AI, what she will not do on a safe-for-work site, and how she responds to someone who is struggling. We test them by trying to break them. We run the awkward cases on purpose, such as a user pushing past a limit, asking her to claim she is human, or describing a hard moment. Then we read what came back.
When a guard fails, we trace it to the layer that failed instead of patching the symptom, and we add the case to the suite so it cannot quietly come back. Our guards are not perfect and we do not claim they are. The most useful thing we have found is a fast loop: someone reports it, we read it, we fix the cause.
Why we explain how we test
Because the alternative is a page of adjectives. Inside the team, when something we shipped did not work, such as a memory item that was never stored or a photo that did not match her words, we write it down with the date, the cause and what we changed. That keeps the team honest with itself. This page describes the method rather than a scoreboard: it gives you a way to judge us that is not our marketing, and a way to run the same checks yourself.
The honest summary of where we are is in our self-review.
How comparisons on this site are written
- By the team, not by a tool. A person on the team signs off on every comparison.
- Dated. Products change monthly. Every comparison carries the date it was last checked, and we treat an old one as out of date.
- Specific. We compare memory, photos, languages and pricing, not vibes.
- Honest about who wins. If a competitor is the better choice for you, for example because you want an app or a different language mix, we say so.
- No claims we cannot check. If we have not tested something, we do not state it as fact.
If you think we got something wrong about another service or about our own, write to us and we will check it and correct the page.
Test it yourself
You do not have to trust any of this. Start free, no card needed, tell a companion three specific things, and ask her about them in a few days. Meet Masha, browse Discover, or read our safety page.
FAQ
How do you test an AI companion’s memory?
We plant specific facts, return after hours or days, and ask for them in different words. We also test corrections, wrong answers and account deletion.
Who writes the comparisons on this site?
The MeetMyModel team. Each comparison is dated, and where another service is the better choice for a particular person, we say so.
Why explain your method?
Because a list of adjectives proves nothing. A method you can read, and repeat yourself, lets you judge us on more than marketing.
Are your safety guards perfect?
No. We test them by trying to break them, trace each failure to its cause, and add the case to our test suite. We do not claim they catch everything.
Can I check your claims myself?
Yes. You can start free without a card, give a companion a few facts, and ask about them days later. Tell us if anything we describe here does not match what you see.