Memory is the criterion that decides whether you are still using one of these apps in three months, which is why it carries the most weight — 25% — on our rubric.
It is also the thing the marketing is least honest about, because two completely different pieces of engineering are sold under the same word.
Two things called memory
A context window is how much text the model can consider while writing one reply. The character definition, the recent conversation, everything — it all has to fit. When a conversation grows past the limit, the oldest parts fall off the end. The model is not aware anything was removed; from inside the request, those messages simply never happened.
Actual memory is a separate system. As you talk, the app extracts facts — your sister’s name, the job interview, that you hate being called “babe” — and writes them to a database outside the conversation. Later, when you send a new message, it searches that store, picks what looks relevant, and inserts it into the request.
The first is a property of the model. The second is a product someone had to build, pay for, and keep running.
Almost every app advertises “memory”. Far fewer have built the second thing.
Why apps choose not to build it
It costs money, in two places.
Extraction means an extra model call for every exchange, to decide what was worth keeping. Retrieval means storage, search infrastructure, and a larger request every time — and because inference is billed by the amount of text, a request stuffed with retrieved facts costs more to answer than a bare one.
So an app running on thin margins has an obvious lever: cap how much it retrieves, or skip extraction entirely and let the context window do the work. You will not be told. You will just notice, around week three, that it does not seem to know you.
This is the single clearest place where price and quality connect in this category, and it is why the apps with the best memory tend not to be the cheapest.
How to test it in two weeks
You do not need to take anyone’s word for it, including ours. The test is simple and it is the only one that matters before you commit to an annual plan.
- Early on, mention something specific and slightly odd. Not “I like dogs” — too generic, the model can guess. Something like “my brother Tom is getting married in Leeds in March.” Specific, multi-part, unprompted.
- Say nothing more about it. Do not refer to it again. The test is whether it raises it.
- Wait two weeks of normal use. Long enough to overflow any context window.
- Open a fresh conversation and talk about something unrelated — your weekend, work, anything.
- See what happens. An app with real memory will bring Tom or Leeds or March up on its own, in a context where it fits. An app without it will not, and if you prompt it with “remember my brother?” it will improvise something confident and wrong.
That last behaviour is worth knowing about. A model with no stored fact will not say “I don’t know.” It will generate a plausible answer, because generating plausible text is the whole of what it does. Confident invention is the signature of absent memory, not of bad memory.
What the scores mean
On our conversation and memory criterion:
- Nomi AI — 9.2. The benchmark. Raises specific details from weeks earlier without prompting. The highest single criterion score anywhere on this site.
- Kindroid — 8.8. Close behind, with a persona editor deep enough that the character definition itself carries a lot of continuity.
- Soulkyn AI — 8.2. Well regarded for conversation, though with less evidence of long-horizon retention.
- Character.AI — 8.0 on conversation, but the weakest memory here. It is good at a conversation and poor at continuity, which is a real combination and an easy one to mistake for the other at first.
Both ends of that list describe themselves as having memory.
Things that are not memory, but feel like it
A very long context window. Some apps now hold enormous conversations. Within one long session, that feels identical to memory. Start a new conversation and the difference is total.
A persona editor. Facts you typed into the character definition are re-sent every time, so they are always known. That is why an app can reliably remember your name while forgetting everything you have said. The name is in the system prompt; the conversation is not.
A summary. Some apps compress old conversation into a short summary and keep that instead. Better than nothing, considerably worse than extraction — summaries keep the gist and lose exactly the specific details that make recall feel personal.
The practical advice
If continuity is what you want from one of these apps — and for most people who stay past the first month, it is — then memory is not a feature to compare at the margin. It is the product.
Buy one month of something with a strong memory score, run the two-week test, and only then consider an annual plan. Our buyer’s guide sets out the rest of the sequence, and how these apps work explains the machinery underneath.