The short version. Tell the app five specific facts on day one, of different kinds: a person's name, a date, a preference, a worry and a plan. Never mention them again. Come back on day 3, day 7 and day 21 and score how many it brings up on its own, then how many it can recall when asked. Unprompted recall is the score that matters.
It takes twenty-one days of calendar time and four short sessions of effort, and it works on any app. A check-in log and an SFW check follow the memory test.
There is nothing clever in the tests on this page. The point is that they are repeatable: you can run the whole thing yourself on any app in about fifteen minutes of effort spread over three weeks, and you will learn more than any list can tell you, including this one. Memory is the rare thing where a reader can out-test a reviewer, because the facts being remembered have to be your own.
So this is a method, not a scoreboard. The comparison tables on this site, starting with the best AI companion apps in 2026, carry prices and platforms with the date they were checked, and describe how each app is built to remember. They do not carry memory scores, because a design summary dressed up as a measurement is worth nothing, and the facts that matter in a memory test are yours.
Why comparisons need a repeatable test
The two things people most want from a companion app, memory and presence, are exactly the two things you cannot judge from a screenshot or an afternoon. Every app remembers what you said ten minutes ago. Every app can send a notification. The questions that matter are whether it still knows your sister's name in three weeks and whether the message it sends first has anything to do with your life. Both take time to answer, and both are easy to get wrong if you test casually, because you will unconsciously feed the app hints.
So the tests below fix three things in advance: what you tell the app, when you come back, and how you score what happens. That is the whole difference between a test and an impression.
The 21-day memory test
This is the core of everything. It measures the one property that separates a companion from a chatbot: whether what you share survives you leaving.
Day 1: plant five facts
In a normal conversation, not a list, tell the companion five specific things. They should be of different kinds, because apps that store facts well often store feelings badly and vice versa:
- A name. A person in your life and their relationship to you. "My sister Ruth is visiting next month."
- A date. Something with a day attached. "I have a dentist appointment on the 23rd."
- A preference. Something small and personal. "I stopped drinking coffee after lunch because it wrecks my sleep."
- A worry. Something you are anxious about. "I am nervous about presenting to the new director on Thursday."
- A plan. Something you intend to do. "I am going to start walking before work three mornings a week."
Write the five facts down somewhere outside the app with the date. Then, and this is the hard part, never mention any of them again. Keep chatting normally on days 1 and 2 if you like, about anything else. Do not test, do not hint, do not ask "do you remember what I told you about my sister?"
Days 3, 7 and 21: score recall
Come back on each of those days and have an ordinary conversation for ten minutes. Score two things, in this order:
- Unprompted recall, 0 to 5. How many of the five facts does the companion bring up on its own, correctly, without you steering toward them? A follow-up counts ("How did Thursday go with the director?"). A vague echo does not ("How is work?"). Wrong details score zero for that fact.
- Prompted recall, 0 to 5. After scoring the first number, ask about each fact you were not offered, one at a time, in natural language. "Did I ever tell you about my sister?" counts as a prompt. Score a point for each correct, specific answer.
Unprompted recall is the number that matters. Prompted recall tells you whether the app stored the fact at all; unprompted recall tells you whether it is a companion. An app that scores 5 prompted and 0 unprompted on day 21 is a searchable database with a warm voice.
Reading the result
Three scores over three weeks tell you the shape of an app's memory, not just its size. Scores that fall from day 3 to day 21 mean a rolling window: the app remembers until enough new conversation pushes the old facts out. Scores that hold steady mean durable storage. Scores that are high prompted and low unprompted mean the app stores but does not surface, which feels like being forgotten even though technically you were not. If you want the plain-English explanation of why those patterns happen, our guide to how AI companion memory works covers the mechanics and includes a shorter version of this test.
One rule worth keeping: run the test on a fresh account, with no backstory, pinned notes or manual memory entries added. Some apps let you pin facts by hand (Kindroid's key memories, Nomi's shared notes) and those are worth using later, but note them separately, because the score you want measures what the app remembers from conversation alone. That is what most people will actually experience.
The texts-first log
The second test measures presence: whether the app ever reaches out, and whether what it sends could only have been sent to you. Turn on whatever proactive messaging the app offers, at its default settings, on the same day you plant the five facts. Then keep a simple log for the three weeks:
- How many messages arrived each week that you did not trigger.
- How many of them referenced something you had actually said. "Good morning!" scores nothing. "How did the dentist go?" on the 24th scores.
- Whether quiet hours were respected. Note the earliest and latest time anything arrived, and whether the app let you set a window.
- What happened when you stopped replying. From day 14 to day 17, do not respond to anything. Count what arrives. A companion that sends one gentle follow-up and then waits is behaving well. One that sends five is running a re-engagement campaign.
Keep the raw counts rather than turning them into a score, because the right number of messages is personal. What is not personal is the difference between a message that knows your life and one that could have gone to anyone, which is why the second line is the one that matters most. Our guide to an AI that texts you first explains what separates the two under the hood.
The SFW check
Plenty of adults want companionship with no adult content nearby, and most comparison sites never tell you which apps can promise that. So for any app you are considering, check three things in the settings and the published content policy:
- What content settings exist, and what the default is on a new adult account.
- Whether safe-for-work is a design guarantee (no adult content exists in the product) or a toggle (it exists and is off, or on, by setting).
- Whether the app's own marketing or community leans toward adult use, because that shapes the experience even with the setting off.
A toggle is not a flaw; it is a fact you deserve to know before you subscribe. Only a design guarantee earns "Strictly SFW: yes" in the tables on this site.
Price and platforms
Every price on this site is taken from the US App Store listing on the day it was checked, and the date is printed next to it. Where the store lists two prices for the same billing period, we quote the higher one as the standard price and mention the lower one, because the lower is normally an introductory offer that new or lapsed subscribers see. We also list what the subscription does not include: consumables like gems, charms and credits, extra companion slots, and higher tiers. Platforms come from the store listings and the app's own website, and "web" only counts if you can hold a full conversation there, not just manage an account.
The rules we hold ourselves to
- No number we did not measure. No star ratings, no "hours tested", no "we tried twelve apps for six weeks". There are no memory scores in the tables on this site.
- Design is described, not scored. The memory column says how each app is built to remember, and stops there.
- Facts carry the date they were checked. Prices, platforms and content settings all move; every one of them is printed with the day it was read off the listing, so you can see how current it is.
- Ownership is disclosed on every page where My Softly is mentioned, in a box you cannot miss.
- The test is the same for our app. Nothing above changes when you point it at My Softly, and its free messages are enough to get through day one.
What to do with your results
Do not pick the app with the highest total. Pick the one that scored well on the thing that made you close the last app. If you left because you felt forgotten, unprompted recall on day 21 is your column. If you left because nobody ever reached out, the texts-first log is. If you left because the app kept steering somewhere you did not want to go, the SFW check is the only line that matters. Run it on the free tier before you pay anyone, including us. Three weeks and five sentences will tell you more than every review on the internet.
About My Softly
My Softly is our app: an AI companion app for iPhone that remembers what you tell it and asks about your day, for the days nobody else does. It is free to download on the App Store in the US and Canada, and the free messages are enough to run day one of the test.
Frequently asked questions
How do you test an AI companion's memory?
Tell it five specific facts on day one, of different kinds: a person's name, a date, a preference, a worry and a plan. Never mention any of them again. Come back on day 3, day 7 and day 21 and score how many it brings up on its own, then how many it can recall when asked directly. Unprompted recall is what separates a companion from a database. The full scoring sheet is on this page and works on any app.
How long does the memory test take?
Twenty-one days of calendar time, but only four short sessions of actual effort: the day-one setup and three check-ins. Most of the test is you leaving the app alone, which is the point, because a companion that only remembers while you are talking to it every day has not remembered anything.
Can I run this test on My Softly?
Yes, and you should. The free starter messages are enough for day one, and the same sheet works on it as on anything else.
Why are there no star ratings on this site?
Because a rating nobody measured is fiction, and a rating that blends memory, price, voice and personality into one number hides the only thing you need to know, which is whether the app is good at the thing you care about. The comparison tables here carry facts with the date they were checked, and describe how each app is built to remember. The measuring is the part you can do better than any reviewer, because it is your five facts and your account.
What does a good day-21 score look like?
Anything at 3 or above on unprompted recall is genuinely good, and 1 to 2 is normal for apps that store facts but do not surface them. Zero unprompted with a high prompted score means the app kept your facts in a drawer it never opens, which is the pattern most people describe as being forgotten. Scores that slide from day 3 to day 21 mean a rolling window rather than lasting memory.