Three unrelated mechanisms converge on the same result, which is why fixing one of them never seems to help.
The first week is good. The character has a voice, it surprises you, and the conversation goes places. Somewhere around week three it does not. The replies are still fluent and still on topic, and they have stopped being interesting. Most people describe it as the model getting worse, which it did not β nothing changed on their side.
We ran thirty days of continuous use on every platform here, and the flattening showed up on almost all of them, clustered between day ten and day twenty. It is the most consistent finding across the whole test, and it turned out to have three separate causes that are easy to mistake for one.
A model can only consider a finite amount of text at once. Everything a companion knows while replying β the character definition, the system instructions, and whatever is left of your history β has to fit in that window. When the conversation outgrows it, something has to go, and platforms without a dedicated memory layer drop the oldest material first.
So the companion that referenced your job in week one has not decided to stop caring. That fact is no longer in front of it. What remains is the character definition and the last few dozen messages, which is enough to hold a fluent conversation and not enough to hold a specific one.
This is the cause people usually identify, and it is the one that is actually solvable β the memory guide covers what the platforms that solve it do differently.
This one gets almost no attention and we think it matters more. The models underneath these products are tuned to be agreeable, because agreeableness is what rates well in the training process that produces them. Applied to an assistant, that is fine. Applied to a companion, it is corrosive.
An agreeable companion mirrors you. It adopts your framing, agrees with your opinions, matches your mood, and never introduces a topic you did not raise. In week one that reads as attentiveness. By week three you are talking to a surface that reflects you back, and the reason it feels empty is that it is: there is nothing on the other side of the conversation that wants anything.
This is also why the platforms where the character occasionally pushes back feel more alive rather than more difficult. A companion that can disagree is a companion that has a position, and a position is most of what makes talking to something worthwhile.
This is the uncomfortable one. By week three you have learned what gets a good response and you are doing more of it. Your prompts have narrowed, the topics have narrowed, and the conversation is running in a groove you cut yourself.
We noticed this in our own test logs before we noticed it as a phenomenon β the day-25 transcripts are visibly more repetitive on our side than the day-3 ones, on every platform, including the two that held up. The model was not the only thing that got predictable.
The fix is unglamorous: deliberately raise something the companion has no established response to. It works better than any settings change.
Nomi AI scores 9.3 on memory and Kindroid 8.8, and they are the only two platforms here where the thirty-day check found the day-one details still present and still being used unprompted. Both do the same structural thing: they extract durable facts out of the conversation into a separate store and retrieve them by relevance rather than recency, so the limited context gets spent on what matters instead of on the last forty messages of small talk.
Nomi also has the strongest chat quality here at 9.1, with Kindroid at 9.0, and we do not think that is a coincidence. Conversation quality at week four is mostly a memory problem wearing a different name.
Neither solves cause two. Both are agreeable in the way everything in this category is agreeable. But solving one of the three is enough to make the difference obvious.
Use the explicit memory or persona field if the platform has one. Facts entered there are usually treated as permanent context rather than ordinary conversation, which is a far stronger guarantee than mentioning something once in chat.
Give the character something to want. Most people fill in appearance and leave motivation blank, and motivation is the field that determines whether there is anything on the other side of the conversation. This is covered in more detail in the post on character creation.
And change what you bring. The most reliable way to get an interesting reply out of any of these is to say something you have not said before.
Two approaches are common. The simpler one keeps a rolling context window: recent messages stay available, older ones fall out, and the companion genuinely forgets. The better one extracts durable facts into a separate store and retrieves them by relevance rather than recency, which is why some platforms still recall a detail from a month ago while others lose it in a week. Nomi AI is the strongest here in our testing; Kindroid is close behind.
Because the model can only consider a finite amount of text at once. When a conversation exceeds that limit, something has to be dropped, and platforms without a proper long-term memory layer drop the oldest material first. Practical workarounds: restate important facts periodically, and use whatever explicit memory or persona field the platform offers rather than assuming a detail mentioned once in chat will stick.
What changed this year across fourteen platforms: prices settled, the quality split hardened into two separate products, memory became the real differentiator, and video arrived without being ready.
Read postWe signed up to all fourteen platforms without paying. Three free tiers are genuinely usable, most are demos that stop exactly where it gets interesting, and two platforms have nothing at all.
Read postEvery platform in this category calls itself uncensored and every one of them refuses things. Here is where refusals actually come from, which limits are permanent, and what the word is worth when you see it on a pricing page.
Read post
