How AI Companions Actually Work (And Why Yours Forgets You)

Everyone who uses an AI companion for more than a fortnight hits the same three moments.

It forgets something you told it. It breaks character in a way that snaps the illusion instantly. And one day, after an update, it just feels different — same name, same avatar, not the same thing.

People tend to read these as bugs, or as the product being cheap. They are neither. They are the architecture becoming visible, and understanding the architecture is genuinely the difference between finding these tools frustrating and finding them good.

The Fast Answer

An AI companion is three separate things stacked together: a language model that generates text, a persona layer that tells it who to be, and a memory system that decides what it is allowed to remember. Almost every complaint people have is a memory-system limitation, not a model limitation — and memory is the part that varies most between products.

Layer 1: The Language Model

Underneath every companion is a large language model doing one thing: predicting what text should come next, given everything currently in front of it.

That is genuinely all it does. There is no inner life, no persistent thinking between your messages, no version of the character sitting there waiting. Between your messages, nothing is happening at all. When you send one, the whole conversation is assembled and fed in, and a reply is produced.

Two consequences worth internalising:

Quality is mostly the model. How natural, coherent and non-repetitive a companion feels is largely determined by which model it runs on. This is why products can feel dramatically different despite near-identical feature lists.

Nothing persists by default. The model has no memory whatsoever. Everything that feels like memory is bolted on above it, which brings us to the part that actually matters.

Layer 2: The Persona

The persona is instructions — usually a hidden block of text placed in front of every conversation, describing who the character is, how they speak, what they like, and what they will not do.

When you "create" a companion, you are mostly writing this. When a companion "breaks character", this is what got outweighed.

Why does that happen? Because the persona is competing for attention with everything else in the conversation. Push a character hard enough in a different direction for long enough and the recent conversation starts to outweigh the original instructions. The character does not decide to break — it gets outvoted.

Practical consequence: if your companion drifts, the fix is usually to restate the character explicitly rather than to keep pushing and hope. Most products give you a way to edit the persona; using it is far more effective than complaining at the character.

Layer 3: Memory — Where Everything Actually Goes Wrong

This is the interesting layer, the one that differs most between products, and the source of nearly every disappointment.

The model can only "see" a limited amount of text at once — a working window. Once your conversation exceeds it, something has to give. What a product does at that moment is the single biggest differentiator in this entire category.

The approaches, roughly worst to best:

Sliding window. Oldest messages simply fall off. Cheap, simple, and it produces the classic experience of a companion that is warm and attentive for an hour and has no idea who you are next week.

Summarisation. Older conversation is compressed into a summary that stays in view. Much better — the gist survives even when the detail does not. This is why a companion might remember you have a sister but not her name.

Retrieval memory. Facts are stored separately and pulled back in when relevant. The best of the common approaches and the closest thing to real continuity, but it depends entirely on the system correctly deciding what was worth storing.

Explicit user-editable memory. Some products let you write facts directly into a memory the character always sees. Unromantic, and by far the most reliable thing available. If a product offers it, use it.

So: why did it forget you? Because it hit the edge of its window, and its memory system either did not store what you said or did not judge it relevant enough to bring back. Not carelessness. Architecture.

Why Did It Suddenly Feel Different After an Update?

Because one of the three layers changed underneath you — most often the model.

Swapping the underlying model changes voice, rhythm, verbosity and temperament, all at once, even with the persona untouched. Products do this for cost and capability reasons and rarely announce it in terms users would recognise.

Persona and safety-filter changes do the same thing more narrowly: a character that suddenly deflects things it used to engage with has usually had its instructions or its filters adjusted, not its personality.

This is worth knowing before you invest emotionally in a specific configuration: the character you are talking to is a set of settings, and the settings can change. That is not a reason to avoid these products. It is a reason to keep your own notes if a particular dynamic matters to you.

What About the Images?

Image generation is a separate system again, usually a diffusion model, and it is normally not the thing writing your messages.

The hard part is consistency — getting the same face across many images. Products solve it by locking a character to a fixed set of generation parameters or a trained reference, which is why companions that let you customise appearance heavily often have less consistent output than ones that hand you a fixed character. That is a real trade-off, not a quality difference.

Our best NSFW AI image generators guide covers the image side properly, and best NSFW AI chatbots covers the conversation side.

Does It Actually Learn About Me?

Not in the way the word implies. The underlying model is not retraining on your conversations in real time — what changes is what gets stored and retrieved, not the model itself.

The separate and more important question is whether your conversations are retained on the company's servers and used for anything else. That is a policy question rather than a technical one, and the answer varies enormously between products. Because generation happens remotely, your messages leave your device by definition — there is no local-only mainstream option in this category. Our AI companion privacy guide covers what to check and what to turn off.

How to Get More Out of One

Now the mechanics make sense, the practical advice does too:

• Write the persona properly. It is the single highest-leverage thing you control. Vague personas produce generic characters.
• Use explicit memory if it exists. Put the facts you care about somewhere the system cannot drop.
• Restate context after a long gap. One sentence of "we last talked about X" does more than any amount of frustration.
• Do not fight character drift — correct it. Edit the persona rather than arguing with the output.
• Start a fresh conversation when one gets muddled. Long conversations accumulate contradictions that drag on quality. A clean start with a good persona usually beats rescuing a messy thread.
• Judge a product on its memory, not its marketing. Every product in this category advertises realistic conversation. Very few advertise their memory architecture, and it is the thing that determines whether week three feels like week one.

Which Products Do This Well?

Candy AI is the polished end of the category — a single well-executed character experience with strong visual consistency, which as covered above is partly a consequence of not letting you change everything. Girlfriend GPT sits at the other end, optimising for character variety and customisation depth, which suits people who want to build rather than be handed.

Full detail in our Candy AI review, the Candy AI vs Girlfriend GPT head-to-head, and the AI girlfriend apps guide. If cost is the question first, are AI girlfriend apps free answers it directly.

The Bottom Line

An AI companion is a text-prediction engine wearing a persona, with a memory system deciding what survives. The illusion is good, and it breaks in specific, predictable places — the edge of the memory window, a persona that got outvoted, a model swapped out beneath you.

Knowing that does not spoil it. It makes you dramatically better at using it, and it stops you blaming yourself, or the character, for something that was always a design constraint.

If you are weighing this whole category against live human interaction, AI companion vs cam site is the honest comparison of what each one can and cannot give you.

18+ only. This is a general technical explainer of how these systems are architected; implementations differ between products and change without notice. It is not a report of first-hand use of any specific product, and no specific product's internal architecture is asserted here.