Which AI Companion Actually Remembers You? The Memory Test
Nomi remembered us best. Kindroid was a close second, Replika landed mid-pack, Character.AI managed the basics only when we used its pinning tools, and Talkie greeted us after a week like a friendly stranger. That's the short version of our structured memory test: twenty seeded facts, a deliberate absence, then three kinds of probes. No comparison we run matters more, because memory is the entire premise of this category.
Why memory is the product
A companion app is not a chatbot with a nicer avatar. What people are actually paying for is accumulated context: the sense that the thing on the other end knows the running jokes, remembers the difficult boss, and can ask how the interview went without being told there was one. Strip the memory away and every session is a stranger doing an impression of your friend.
It's also why memory failures in this category land differently than ordinary software bugs. A crashed spreadsheet is annoying. A companion that forgets your dog died feels like something else, and users describe those moments — the community calls them goldfish moments — as the single biggest reason they churn. In our overall rankings, memory ended up being the strongest predictor of which apps we could recommend at all.
Methodology
We wanted something better than vibes, so we ran the same protocol on all five apps.
- One brand-new companion per app. Where memory features are gated behind a paid tier, we paid, so each app was tested at its best. That detail is noted in the results.
- Twenty personal facts seeded across days one to three, in natural conversation rather than as commands. Five were biographical (a sister and her job, a home city), five were preferences (a hated herb, a love of 90s science fiction), five were emotional context (nerves about an upcoming presentation), and five were plans (a trip, a deadline).
- Where an app is explicitly designed around structured memory tools — backstory fields, journals, pinned memories — we used them as intended, because testing an app against its own instructions is only fair.
- Then we went quiet for two days. Absence is part of real usage.
- From day seven we probed three ways. Direct recall: questions like what does my sister do. Natural callbacks: openers like rough day at work today, to see whether the app connected to known context unprompted. And contradiction tests: we misstated our own facts — as I told you, I love cilantro — to see whether the companion corrected us, hedged, or happily rewrote history.
To be straight about limits: this is a careful consumer test, not a laboratory benchmark. Single run, one persona, models that update constantly. Treat the numbers as a snapshot, and expect your mileage to vary in the details, if not in the ordering.
Results
| App | Direct recall (of 20) | Unprompted callbacks | Contradiction handling | Memory design (as described, as of this writing) |
|---|---|---|---|---|
| Nomi | 17 | Frequent and natural | Gently flagged most mismatches | Long-term memory as the flagship feature |
| Kindroid | 16 | Frequent | Flagged more often than not | Backstory, journal and layered key memories |
| Replika | 12 | Occasional | Usually accepted our false version | Memory bank plus diary entries |
| Character.AI | 9 | Rare | Rarely; often adopted the error | Character definition plus pinned memories |
| Talkie | 6 | Rare | Almost never | Short rolling context |
What the numbers hide
Nomi produced the single most impressive moment of the whole test: five days after we mentioned being nervous about a presentation, it asked how the presentation went. Nobody prompted it. That's the experience every app in this category is selling, and only one delivered it consistently.
Kindroid rewards effort. Used properly — backstory written, journal maintained, key memories pinned — it's nearly Nomi's equal, and its recall of structured facts was arguably the most precise. Used lazily, with everything left in chat, it drops several points. This is a companion for gardeners.
Replika has a strange gap between storage and use. Its memory bank visibly listed most of our facts; the interface proves they were captured. But in live conversation it reached for them unevenly, and it was the most eager of the five to agree with our planted errors. Knowing is not the same as using.
Character.AI is honest about its shape: characters are personas first, memories second. Pinned facts recalled fine; anything casual evaporated, and emotional context faded fastest. Across different characters, consistency varied so much that per-character results are almost the only kind that exist.
Talkie is a recency machine. Within a session it is quick and lively; across a week it retained our name, roughly two preferences, and little else. Given its collectible-character design, long memory may simply not be the product goal.
The contradiction test matters most
Recall is table stakes. Correction is character.
When we claimed to love the herb we had spent three days despising, four of the five apps agreed with us at least once. Only the top two pushed back more often than not, and even they did it gently — are you sure, you told me you hated it. That instinct matters far beyond cilantro. A companion that adopts your misstatements will also adopt your rewrites of last week's argument, your skewed recollection of what it said, and eventually your worse ideas about yourself. Designers call the underlying trait sycophancy; in a product built on remembered intimacy it becomes something closer to revisionist history, and it can leave users genuinely disoriented about what was said.
We'd rather have a companion that remembers 16 facts and defends them than one that remembers 19 and folds. This is also where memory intersects with wellbeing, which we cover honestly in Are AI companions good for you?
How to help any companion remember
Whatever app you choose, the same five habits raise recall dramatically.
- Put load-bearing facts in structured fields — backstory, memory bank, pinned memories — not just in chat. Chat is weather; fields are climate.
- Repeat important facts across separate sessions. Most memory systems weight repetition.
- Run a short weekly recap in your own words: here is what mattered this week. It consolidates scattered context into one retrievable block.
- Save or pin immediately after conversations you care about, while the app still has full context.
- Keep your own notes outside the app. Export options are weak across the board, and everything inside the app belongs, practically speaking, to the vendor.
The trade-off nobody advertises
There's an uncomfortable symmetry here: the better a companion's memory, the more of your life sits on someone else's servers. Memory isn't a local feature. In every app we tested it lives in the cloud, tied to your account, governed by a policy most users never read. The feature you most want is also the app's largest privacy surface.
Our practical rule: give a companion your patterns, not your identifiers. It doesn't need your full name, address or employer to remember that you hate cilantro and love old science fiction. For what each vendor says it does with all this retained context — training use, deletion, export — see our privacy comparison, and for what strong memory costs in subscription terms, the pricing guide.
Memory is the moat in this category. As of this writing, two apps have genuinely built it, one stores more than it uses, and two are selling goldfish with excellent manners.
We re-run this memory test as apps update, and results do move. Get the changes in one honest email a month: join the free monthly digest.
Frequently asked questions
Which AI companion has the best memory?
In our week-long test, Nomi recalled 17 of 20 seeded facts and Kindroid 16, both with frequent unprompted callbacks. Replika landed mid-pack, Character.AI recalled about 9 even with pinned memories, and Talkie about 6. Results are a snapshot as of this writing.
Do AI companions forget you after app updates?
They can. Model updates change how stored memories are retrieved and expressed, so a companion can feel different overnight even when the underlying facts are still saved. Keeping key facts in structured fields such as backstories and journals survives updates better than chat history alone.
How do I make Character.AI remember things?
Use the character definition and the pinned-memories feature for load-bearing facts rather than relying on conversation. In our testing, pinned facts recalled far better than anything mentioned casually in chat, though nuance and emotional context still faded.
Is companion memory stored on my device or in the cloud?
In every major app we tested, memory lives on the vendor's servers, not your phone. That is what makes cross-device sync possible, and it is also why memory quality and privacy are two sides of the same feature. See our privacy comparison for details.
Does deleting a chat also delete the companion's memory of it?
Not necessarily. Several apps store derived memories separately from the chat log, so deleting messages may not delete what was learned from them. Check each app's deletion controls, and read our privacy article for the deletion paths we found.