AI girlfriend app memory: what free tiers keep

It greeted you by name on Tuesday, so you assumed it was keeping up. On Wednesday you mentioned the trip again and got a polite question about where you were going. AI girlfriend app memory is not one feature that either works or does not — it is three separate mechanisms advertised with the same word, and a free tier usually gives you the cheapest of them. Knowing which one you are looking at is the difference between judging an app fairly and paying to fix a problem it does not have.

Three things apps call memory

These are not marketing distinctions. They are different pieces of engineering, they cost the company different amounts, and they fail in ways you can tell apart.

• The context window. Every reply is generated from a block of recent conversation handed to the model with your latest message. That block is finite by design, and when it fills, the oldest messages fall out. This is not memory at all, but it looks like memory for as long as a conversation is short. It comes free with the technology, so every app has it.
• A rolling summary. Rather than let old messages drop, the app periodically has a model condense them into a paragraph and carries the paragraph forward. Cheap, and it explains a specific oddity: the character remembers the gist of last week and none of the wording, because the wording is gone.
• Stored facts, fetched back. The app extracts individual details — your name, your job, the dog, the trip — writes them to a database, and pulls the relevant ones into the prompt when they come up. This is the expensive one to build and run, and the one people mean when they say an app remembers them.

So a character can remember your name for months while losing a conversation from two days ago. The name was stored as a fact; the conversation was only ever in a window.

What a free tier usually keeps

Free tiers economise here harder than anywhere else: storing and retrieving facts for every user who never pays is a cost with no ceiling. The shapes repeat across the category.

• Session-only. Close the app and the conversation starts clean. Often paired with generous message counts, since messages are cheap when nothing is kept.
• A short window, no storage. The character holds the last stretch of the conversation and nothing before it. The tell is one that contradicts something you agreed twenty minutes ago.
• A small fixed number of stored facts. Real storage, capped low enough that new facts push old ones out. Some apps show the count.

None of this appears on the plan page, which says "memory" with a tick or a cross. The faster route is to measure it.

Testing AI girlfriend app memory in one evening and one morning

You can separate the three mechanisms in about twenty minutes of typing, split across a night's sleep. Run it on both shortlisted apps at once.

1. Plant two facts, far apart. Early on, mention something specific and checkable — a pet's name, a city you are visiting on Friday. Then talk about something else for as long as you can stand.
2. Ask about the first fact before you leave. If it is gone already, you are inside a short context window and there is no storage behind it. That is the whole answer, and it took one sitting.
3. Close the app and come back tomorrow. There is no way to run this one faster than it elapses.
4. Open a new conversation rather than the old one. Many apps keep a thread alive but store nothing centrally, so continuing yesterday's chat flatters them. A fresh conversation is where stored facts show up or do not.
5. Ask about the detail, not the gist. "What did I say I was doing Friday" separates stored facts from a summary. A summary answers with the theme and misses the day.
6. Correct it once, then check the correction held. Tell it the trip moved to Saturday, then ask again later. Storing corrections is harder than storing first impressions, and not every app does it.

Record what each app failed at rather than whether it felt attentive. Our method for comparing free tiers puts this beside the message cap and the voice allowance in one week, and this is the row where two apps diverge most.

The feature to look for is a memory you can read

The most useful thing an app can give you here is not a longer memory but a visible one: a panel listing what it has stored about you, where you can edit or delete a line. Some apps put it on the free tier, and it turns the testing above from guesswork into reading.

It also exposes two things worth knowing before you pay. Memory is usually per character, so a detail you told one companion is unknown to another and a second character starts from nothing. And on many apps, editing a character's personality settings does not clear what has been stored about you, which is how a rewritten character still brings up a fact you thought you had left behind. Our note on what the personality controls actually change covers where that line falls.

It is also where you find out that extracted facts drift. A passing remark becomes a preference, a joke becomes a biographical detail, and the character keeps returning to it because a database says it matters.

Memory you can remove

Stored facts are stored data, which puts them under whatever privacy regime covers you rather than under the app's goodwill. In the UK and the EU that includes a route to ask for your data to be deleted.

Three things to check before you invest a month of conversation: whether a single stored fact can be deleted in the app, whether deleting a character removes what it stored about you or only hides the chat, and whether "clear memory" also clears the summary behind it. The privacy policy settles this more reliably than the settings screen; our privacy checklist lists the sections to read.

Where a long memory works against you

Here is the part the upsell leaves out: more memory is not uniformly better. A system that stores a great deal and retrieves eagerly produces a character that keeps steering back to the same handful of facts, because those facts are what it has. That is often what people mean when an app turns repetitive in month two, and the cause is the feature they paid for.

Stored facts also go stale in a way a window never does. A preference you mentioned once outlives the truth of it, and unless you can find and edit the note, the only fix is correcting a character that is working confidently from an old file.

So the memory worth paying for is the one you can inspect and correct, not the one with the largest number beside it. To run the two-day test without handing over a card, the app we currently recommend keeps stored facts on its free tier and shows you the list, which is what makes the test readable rather than inferred.

Frequently asked questions

Does AI girlfriend app memory work on a free tier?

Something usually does, but most often it is the recent conversation rather than anything stored. Close the app, come back the next day and start a fresh conversation: if a specific detail survives that, the app is storing facts rather than holding a window open.

Why does the character forget what I said yesterday but remember my name?

Because those are handled by different mechanisms. A name is typically extracted and stored as a fact, while a conversation lives in a context window that empties as it fills, so the detail goes and the name stays.

Can I delete what an app has remembered about me?

Many apps let you clear stored memory, and some let you delete individual entries. Whether that also removes the underlying chat logs is a separate question answered in the privacy policy, and in the UK and EU you have a route to request deletion directly from the company.