Context windows versus retrieval-based memory, why most companions lose track of you within a fortnight, and what the platforms that get it right are doing differently.
Last updated:
A language model can only consider a finite amount of text at once. Everything a companion 'knows' during a reply has to fit inside that window: the character definition, any system instructions, and as much conversation history as remains.
When a conversation outgrows the window, something has to go. Platforms without a dedicated memory layer drop the oldest material first, which is why a companion can be warm and consistent for a fortnight and then quietly stop remembering that you have a sister.
The stronger approach extracts durable facts from the conversation into a separate store, then retrieves them by relevance rather than by recency. Asked about your job, the system pulls the stored fact about your job regardless of how long ago you mentioned it, and spends its limited context on that rather than on the last forty messages of small talk.
This is the difference our 30-day memory protocol is designed to expose. We plant specific facts on day one and check recall at day 7 and day 30 without prompting. Nomi AI was the only platform in the comparison that consistently passed the day-30 check β and passed it well, bringing facts up naturally when relevant rather than only retrieving them on request. Kindroid finished second and held most day-one details. Everything else degraded noticeably.
The same identity problem shows up visually. Most image generation is prompt-to-image with no persistent identity, so each generation is independent and faces drift between them.
Companion platforms that tie generations to a stored character definition do markedly better β Candy AI holds a character across sessions notably well. Dedicated generators like SoulGen and PornPen produce stronger individual images but are weaker at keeping one character recognisable over time, which is the trade-off to weigh if you want a recurring character rather than a lot of pictures.
Use the explicit memory or persona field if the platform has one. Facts entered there are usually treated as permanent context rather than as ordinary conversation, which is a much stronger guarantee than mentioning something once in chat.
Restate important details periodically. It feels artificial, but on platforms without a real memory layer it is the only thing that works.
Be sceptical of 'long-term memory' on a feature list. Nearly every platform claims it; the claim is close to meaningless without testing across weeks, which is why it is scored separately here.
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.
Because most image generation is prompt-to-image with no persistent identity β each generation is independent, so faces drift. Companion platforms that tie generations to a stored character definition do much better at this: Candy AI holds a character across sessions notably well. Dedicated generators like SoulGen and PornPen are stronger per-image but weaker at keeping one character recognisable over time.
The same generator is almost never good at both styles, and picking by style narrows fourteen platforms to two or three faster than any score does. What each end of the category is actually good at.
Read postSix of fourteen platforms generate video. What you actually get is three to six seconds, no audio, and a token cost that makes it the most expensive thing on the menu. Where it is worth it and where it is not.
Read postThe flattening is real, it happens on almost every platform, and it has three separate causes that people usually blame on one. What we found across thirty-day tests, and the two platforms it does not happen to.
Read postAppearance and personality are two different systems behind one form, they behave completely differently, and almost everyone spends their time on the one that matters less.
Read post
