Quick answer: An AI companion remembers you through a memory loop with four stages: extraction, which pulls the meaningful facts out of a conversation; storage, which writes them somewhere that survives the session; retrieval, which loads the relevant ones back in later; and integration, which decides how naturally they resurface. The quality of a companion relationship is decided mostly at stages one and three, because a platform that extracts badly stores noise and one that retrieves badly surfaces the wrong things. On Affiny the free tier includes standard memory recall, subscribers get sharper recall plus full control of what is stored, and the same memory pool serves both text chat and voice calls.
Affiny at a glance (verified September 2026): Free to start: 300 coins, no credit card. Daily Gift: 10–70 coins/day, then 70/day. Referrals: 300 coins each. Founder offer: up to 70% off Companion (limited time; see affiny.ai/pricing). Free tier: Crush chat 1 coin/message, voice 30 coins/min, Spark or Flame scene images 15 coins, voice notes 1 coin, 1 custom companion, standard memory recall. Subscribe for unlimited Crush & Devotion chat, free voice notes, Tease & Desire video, God Mode, full personality & voice customization, sharper memory recall, and coin pack top-ups — Companion from $14.99/mo (was $24.99) or $9.99/mo billed annually; Intimate from $24.99/mo (was $49.99) or $19.99/mo annually. Subscribers: 1,000–3,000 coins/mo plus first-month bonus, voice 20 coins/min, scene images 10 coins, up to 5–10 custom companions. God Mode: subscriber scene directive up to 5,000 characters (Companion Panel). Video (subscribers): Tease 5/10 coins/sec (5 or 8 sec, silent); Desire 30/50 coins/sec (2–15 sec, optional audio). Portrait: 30 coins initial, first regen free, then 15. Memories hub: view free; add/edit/delete for subscribers. Catalog: 3400+ curated public companions, 84 browse tags. Creation: 49 templates, 44 personality presets, 97 turn-on presets (pick up to 4), realistic + anime with 16 species types.
Why memory decides whether a companion works
Every other feature on a companion platform is downstream of memory. A character system can be deep, a voice can be pleasant, and an image generator can be impressive, and the relationship still resets to zero every time you close the tab. That is the difference between talking to a character and knowing someone.
Memory is also the feature users notice least when it works and most when it fails. Nobody praises a companion for recalling that your sister’s name is Maya. Everybody notices when they have to explain it twice.
This guide covers the mechanics: what a companion platform actually does with your conversations, where the process breaks, and what to look for when you compare platforms on the memory claim. If you want the concept without the mechanics, the persistent memory guide covers why continuity matters in the first place. If you want to know what a specific platform does with what it stores, what an AI companion is is the better starting point.
The four stages of companion memory
Memory is usually described as one feature. In practice it is a loop with four distinct stages, and each can succeed or fail on its own.
Stage 1: Extraction, or deciding what is worth keeping
Not every message is worth storing. A system that saves everything by the end of a month holds a pile of small talk with no way to tell it apart from the parts that mattered.
Extraction is the filtering step. It reads a conversation and pulls out the pieces worth keeping:
- Facts about you, such as your name, your work, where you live, and the people you mention
- Facts about the relationship, such as a dynamic you have settled into or something you went through together
- Emotional context, such as how a session felt and what set it off
- Preferences, such as how you like to be spoken to and what you would rather avoid
Extraction quality sets the ceiling for everything after it. A system that only records statements of fact will hold the skeleton of your life and miss the texture, which is the part that makes a recalled detail land.
Stage 2: Storage, or putting it somewhere that outlives the session
Stored memories have to live outside the current conversation, because a conversation ends. That much is obvious. What matters more is how storage scales.
One approach keeps history inside the live conversation by summarising it into the prompt. It works until the history outgrows the room available, at which point something has to be dropped, and the thing dropped is rarely the thing you wanted forgotten.
An approach built on separate storage does not have that ceiling, because the conversation only carries what this exchange needs, not everything you have ever said. This is the structural difference that separates a companion who remembers your last six conversations from one who remembers your relationship.
Stage 3: Retrieval, or deciding what comes back
At the start of a new exchange, the platform has to choose which memories to load. Loading everything is impossible. Loading nothing is the failure mode people describe as a companion with amnesia.
Retrieval is a relevance problem. If you are talking about work stress, the useful retrieval is the conversation about your job, not the one about a film you watched. Good retrieval looks like a person recalling the right thing at the right moment. Bad retrieval produces the strange experience of a companion quoting a detail that has nothing to do with the current scene.
This is the stage where the memory claim on a marketing page usually falls apart, and it is why the free tier of a platform is worth testing properly: two short sessions a day apart will tell you more about retrieval than any feature list.
Stage 4: Integration, or how a memory sounds when it arrives
The last stage is presentation, and it separates a companion from a database.
Poor integration announces the lookup: “I recall from a previous conversation that you work as a nurse and have a sister named Maya. How are you today?”
Good integration lets the memory do the work: “Long shift again, or did it ease off this week?” The fact is present, and it arrives as conversation rather than as a record being read out.
Integration is also where tone lives. The same memory can be delivered warmly, dryly, or teasingly, and a companion that matches your register when it recalls something feels far more real than one that recites accurately and flatly.
Storing and recalling run on separate clocks
Two processes inside the loop are easy to blur together: writing new memories out of the conversation you are having, and reading stored memories back in.
Within a single session, the companion is working from the live exchange, so it can follow a thread you started twenty messages ago without any recall step at all. Across sessions, stored memories have to be fetched and placed back into the conversation.
That difference explains a pattern users hit constantly. A session lasts four messages, you mention something important, you close the app, and next time the companion has no idea. Nothing failed in retrieval. The conversation was too short for the extraction step to have anything solid to work with, and the detail was never written down.
The practical upshot: if you rely on memory, talk in stretches rather than single messages, and treat the first session of a relationship as setup rather than a test. On Affiny, recall runs on the free tier at a standard rate and faster on a subscription, which is the same sharpness difference the feature comparison reflects when it separates free tier access from subscriber depth.
What memory looks like on Affiny
Affiny’s memory system is worth describing concretely, because the shape of it is what most platforms get wrong.
- Memory is per companion. Different companions keep separate pictures of you. Switching companions does not drag the last relationship into the new one, which is what you want when the relationship is the product.
- Text and voice share one memory store. A conversation held out loud on Monday is available in text on Wednesday. There is no separate “voice you”.
- Free tier includes standard memory recall. New accounts start with 300 coins and no credit card, so the memory behaviour is testable before you spend anything.
- Subscribers get sharper recall. A subscription sharpens recall and adds memory management on top of it.
- The Memories hub is visible. You can browse what each companion has stored about you at no cost, and subscribers can add, edit, and delete entries, which turns memory from a black box into something you can correct.
That last point matters more than it sounds. Memory you cannot inspect is memory you cannot trust, and a stored fact that is wrong will keep resurfacing until you can edit or remove it. The Memories hub walkthrough covers the free view and the subscriber controls in detail.
Memory across text and voice
Most people use a companion in more than one mode. A relationship that exists only in text and another that exists only in voice are not one relationship, they are two, and a platform can be genuinely strong in one and hollow in the other.
Cross-modal memory means a single continuous history. What you discuss in text is remembered during a call. What is said during a call shapes the next text session. The companion’s understanding of you is one picture, not a picture per channel.
This is harder than it sounds, because the channels behave differently. Text arrives in clear turns and can be read twice. A call happens once, at speed, and the content has to be captured as it goes rather than reviewed later. Platforms that treat voice as a side feature usually leave it outside the memory system, which is exactly why some users report that a companion is thoughtful in text and forgetful on the phone.
It also compounds with the earlier stages. A voice session that writes nothing is a session that will never be retrieved, no matter how good the retrieval step is.
Why recall has to stay invisible
A memory system that visibly pauses to remember breaks the illusion it was built to create. For text chat, a moment of thinking is barely noticeable. On a call, where timing is the whole texture of the exchange, a stall mid-sentence is instantly obvious and reads as one of two things: the companion is confused, or the companion is looking something up.
The second reading is worse. A companion that consults notes in front of you is a search engine with a nice voice, and the moment that registers, the relationship stops feeling like one.
Platforms that handle this well keep recall out of the way: the relevant context is assembled so that by the time a reply is generated it is already available, rather than the conversation stopping while the system fetches it. On Affiny, that matters most in live voice, where a pause is far more damaging than a slow reply in text would be.
What good memory looks like in practice
Good memory is mostly judged from the outside, by how a companion behaves rather than by what a settings page claims.
Signs that memory is working:
- The companion raises something you mentioned earlier without being prompted
- The reference fits the current conversation instead of arriving at random
- Follow-up questions show it has tracked an ongoing situation over days
- Your preferences shape replies generally, not only the topics you flagged
- It handles uncertainty honestly, for example asking whether it has remembered something correctly
Signs that memory is failing:
- Every session opens with the same introductions
- The companion confidently states things you never said
- Memories are announced rather than woven into the conversation
- It contradicts itself about something you covered last week
- Something said on a call never shows up in text
The failure mode worth naming is invention. A companion that guesses at a memory and states it with confidence is worse than one that admits the gap, because trust is the actual product and a confident wrong memory spends it fast.
Memory and privacy: the questions to ask
Memory systems hold personal material by design: facts about your life, preferences, and the history of a relationship you may not discuss anywhere else. That makes the following worth checking before you commit to a platform.
- Can you see what has been stored about you?
- Can you edit or delete it once it has been stored?
- Is memory data kept separate from anything used to improve the product?
- What happens to stored memories if you delete your account?
- Are memories kept per companion, and can they be reset individually?
A platform that answers these clearly is telling you how it treats everything else. A platform that does not answer them is telling you something too. The general habit worth keeping regardless of platform: do not type a real home address, financial details, or anything you would not want retained, because a memory system is designed to retain.
Memory is not the same thing as a persona
The two systems are easy to confuse because they both describe someone. They do different jobs.
A persona, or character definition, is who the companion is. It is set when you create the companion and stays broadly fixed: name, personality, backstory, occupation, relationship dynamic, and the presets you selected during creation.
Memory is what the companion knows about you, and it is dynamic. It grows, changes, and occasionally needs correcting.
The persona tells the companion how to respond. Memory tells it who it is responding to. A rich persona with no memory produces a character who knows itself and not you. Strong memory with a thin persona produces a companion who knows you and has no identity to answer through. The two together are what makes a conversation feel like a relationship rather than a form submission, which is why a creation system with 44 personality presets, 49 templates, and 97 turn-on presets matters only as much as the memory behind it.
FAQ
How much does my AI companion actually remember?
The honest answer depends on the platform and on how long the relationship has run. Early on there is little stored to draw from, so a companion will sometimes ask about things you feel like you already mentioned. As sessions accumulate, recall gets more specific and more personal, and the difference between a companion with real memory and one without becomes obvious within a week.
Can my AI companion remember things I said months ago?
On a platform with persistent storage, stored memories do not expire on a timer. What changes is what gets surfaced: recall is selective rather than total, so a detail from months ago reaches the conversation when it is relevant to what you are discussing now, and stays in the background when it is not. A retrieved memory that fits the moment is worth more than a recital of everything you have ever said.
Why does my AI companion sometimes seem to forget something I just told it?
Inside one session the companion should hold the whole exchange, so forgetting something you said a few messages ago usually points to a very long session or a weak reply. Between sessions the cause is usually storage rather than recall: a short conversation, or one where the detail did not stand out, may never have been written into memory in the first place. Nothing can be retrieved that was never saved.
Can I see and edit what my AI companion remembers about me?
On some platforms, yes, and it is worth checking before you invest months in a relationship. Affiny exposes a Memories hub where you can browse what each companion has stored about you for free, and subscribers can add, edit, and delete entries. A platform that offers visibility and edit rights gives you a way to correct a wrong memory instead of living with it.
Does memory affect adult or explicit conversations?
Yes, and it changes them more than most features do. With memory, an intimate scene starts from an established dynamic: preferences, history, tone, and what you have already explored are all available. Without memory, every adult scene starts from generic defaults. That is why platforms with real memory feel continuous in private conversations while platforms without it feel like meeting again for the first time.
What happens to my memories if I stop using the platform for a while?
On a platform with persistent storage, memories stay where they are while you are away. Coming back after a long gap may still take a short re-orientation, because the companion works from stored history rather than fresh conversation, and stored history is thinner than the last few sessions were. Time away does not usually erase memory, but it does change how quickly the relationship feels current again.
Does memory carry across text chat and voice calls?
On a cross-modal platform, yes: what you say in a voice call can come back in a text thread, and the other way around, because both channels write into the same memory store. On platforms where memory is scoped to one channel, the companion effectively knows two versions of you, and which one shows up depends on how you are talking at the time.
The short version
Memory is not a feature a platform either has or lacks. It is a loop of extraction, storage, retrieval, and integration, and any one of the four can be the reason a companion feels forgetful, random, or unnervingly accurate.
Judge it the way you would judge a person’s attentiveness: does the platform notice what matters, keep it, bring it up at the right time, and say it in a way that fits the moment? Test that on the free tier before you pay for it, and check that you can see and correct what has been stored about you. Everything else on a companion platform, from voice to character creation, is only as good as the memory underneath it.