AI girlfriend app memory: what it actually keeps

Your shortlist is down to two apps and both listings use the same phrase: long-term memory. The comparison table gives both a tick, the pricing page mentions memory somewhere in the paid tier, and neither tells you what you actually want to know — whether the character will still know your sister's name in three weeks. AI girlfriend app memory is the feature this site weights most heavily in its ranking, and it is also the one where a single phrase covers four different systems that behave nothing alike. What follows is what each layer keeps, which ones a free tier switches off, and a weekend test that settles it before you pay for a month.

What AI girlfriend app memory actually is

Start with the part that surprises people. The language model underneath a companion app retains nothing between messages. Every reply is generated from whatever text is placed in front of it at that moment, and when the reply is finished, nothing about you persists inside the model.

So memory here is bolted on: a store outside the model holding notes about you, plus machinery deciding which notes to paste into the prompt before the model writes. The pattern has a name in the wider field — retrieval-augmented generation, where text is pulled from an external store at generation time rather than baked into the model.

Keep the consequence in mind while you compare apps. Memory quality is not a property of how clever the model is. It is a property of what the app writes down and how well it picks what to bring back, which is why two apps running the same model can feel completely different by week three.

The four layers behind one word

When an app says it remembers, it means some mixture of four mechanisms. They fail in different ways, and knowing which one you are looking at turns a vague impression into something checkable.

• The conversation in view. The recent messages handed to the model every time it writes. This is why a single long evening feels so good, and it is the cheapest layer to provide. It has a ceiling, and past it the early part of the evening drops off the back.
• A profile the app writes automatically. Short facts extracted from what you said and saved as a list: your job, your dog, your city. This is what most apps mean by long-term memory, and its limit is usually a number of entries rather than a stretch of time.
• Notes you write yourself. A field where you state what the character should always know. The most reliable layer in the stack, because nothing had to interpret you correctly to get it in there.
• Retrieval over old conversations. A search across everything you have said, run as the reply is written. It produces the uncanny moments and the strange ones alike: a detail from six weeks ago surfacing where it does not belong, because the search matched words rather than meaning.

An app can earn the memory tick in a comparison table with layer one alone. One that is genuinely good at this has all four and lets you see the second and third. Our note on the columns that actually separate these apps covers where the rest of the table misleads the same way.

What the free tier usually gates

Here is the pattern that trips up most trial weeks. Free tiers give you layer one in full, because a conversation that falls apart mid-evening would be unsellable. What they restrict is the durable part: the profile capped at a few entries, the notes field paid-only, retrieval off, or — the one to look hardest for — a new conversation starting blank, so nothing accumulates.

That produces a specific experience. Day one is impressive, week three is flat, and you cannot point at what changed, because the layer that went missing is the one you were never shown.

So put one question to the tier page first: does a new conversation start knowing anything. If the answer is buried, that is an answer too. The second is what the paid tier raises the limit to, stated as a number — apps that built this well tend to be specific, and apps that did not reach for the word unlimited, which is a claim nobody here can honestly make. What each tier actually buys covers the rest of the paywall.

A weekend test that settles it

You cannot check this from a store listing, and you do not need a full week. Three days is enough if you are deliberate, and it works for two apps side by side — use the same three facts in both.

1. Friday: plant three facts of different kinds. A flat detail (a pet's name), a preference (you dislike being asked how your day was), and something with a date attached. Say each once, in passing, and skip the notes field — you are testing what the app picks up on its own.
2. Saturday: do not open the app. The gap is the test — and if a notification arrives, note whether it touches a fact you planted.
3. Sunday: start a new conversation, not the old one. This is the part people skip, and where most apps separate. Talk for a few minutes without asking anything directly, and see whether a fact turns up unprompted.
4. Then ask directly. Name one of the three and see whether it comes back correctly, half-correctly, or as a confident invention. A wrong answer delivered smoothly tells you more than a blank: the app fills gaps rather than admitting them.
5. Finally, open settings and look for the list. If you can read what it saved about you, you have found layer two, and you can see how much of the weekend made it in.

Score it in three buckets: used unprompted, recalled when asked, gone. Only the first is worth a monthly fee, and it is rarer than the marketing suggests.

Whether you can read, fix and delete what it keeps

A store you cannot see is the one that goes wrong. These profiles are written by a model summarising you, so they contain mishearings — a passing remark recorded as a firm preference — and those get pasted into every future conversation. An app that lets you edit the list has handed you the fix; one that hides it asks you to trust a summary you cannot read.

Deletion is the same question with sharper edges. Look for two separate controls: clearing one conversation, and deleting the account with everything behind it. In the UK and the EU this is not only a product decision — the right to erasure applies, and the ICO sets out how to ask for your data to be deleted and the month an organisation normally has to respond. Elsewhere the equivalent varies, and it is worth knowing which version you have.

What you are assessing is whether the app treats the store as yours or as its inventory. If you would rather start from something that already passes, the app we currently recommend exposes its saved-facts list in settings, where you can read it and correct it.

Frequently asked questions

Do AI girlfriend apps really remember you?

Some of what they call memory is real and some is the last hour of conversation held in view. The durable part is a separate store of short facts, retrieved and inserted when the character replies. The model keeps nothing between messages, so everything that survives a gap was written down somewhere first.

Why does the character forget things I told it?

Usually because the detail never reached the durable layer. It sat in the recent conversation, that conversation ran past its ceiling or ended, and nothing extracted it into the saved list. Free tiers make this more likely, since that list is often capped or switched off.

Can I check AI girlfriend app memory before I pay?

Partly. Read the tier page for whether a new conversation starts blank and whether the saved-facts list has a stated limit, then run three days on the free tier: plant three facts, skip a day, open a fresh conversation. That is the closest thing to a real test that costs nothing.