What Memory Means When a Companion App Claims It
A claim of memory can describe at least four different behaviours, and they are not equivalent. It can mean the app holds one conversation together while it is happening, that specific facts carry from one conversation into the next, that there is a list of things you explicitly asked it to keep, or that a running summary is maintained and quietly rewritten as it grows. The marketing word is identical in all four cases. Ten minutes of deliberate testing tells them apart.
Doing that testing early matters, because the variety you are getting determines what happens when you come back after a fortnight away — and by then you have invested the conversations that would have been the evidence.
The four behaviours the word covers
Holding one conversation together. Everything you have said in this session stays available while the session lasts. This is the weakest form and the one most likely to be described as memory, because from the inside it feels like being remembered. It ends when the conversation does, and on some apps it degrades before that — a long enough exchange starts losing its own beginning.
Carrying facts between conversations. You mention something on Tuesday and it is still true on Friday, in a new conversation, without you repeating it. This is the behaviour most people mean when they search for an app with memory, and it is the one worth confirming rather than assuming.
An explicit list you control. Some apps expose the things they have kept as a readable, editable list — you can add to it, correct it, and delete from it. This is by far the most informative version, not because it is the most sophisticated but because it is the only one where you can see the state instead of inferring it.
A rolling summary. Rather than keep what you said, the app keeps a condensed account of it, and condenses again as the account grows. This behaves like memory for a while and then behaves like a game of telephone: details survive as their gist, then as the wrong gist. The characteristic symptom is an app that remembers roughly who you are and gets a specific fact confidently wrong.
Most apps combine several of these, which is why a single test result rarely settles the question.
Testing which one you have
Plant a fact and leave it. Early in a conversation, state something specific, arbitrary and easy to check — a place, a date, a name of something that is not you. Do not refer to it again. Come back days later, in a new conversation, and ask. This one test separates the first behaviour from the second.
Contradict yourself on purpose. Tell it something, then much later tell it the opposite, then ask. An app that keeps a list will usually hold both or take the newer one; an app running summaries tends to blend them into something you never said. What you learn is less about accuracy than about how the state is being maintained.
Look for a screen, not a reply. If there is a settings page showing what the app has retained, read it. What it lists is the honest inventory. What it omits is either not kept or not shown to you, and both are worth knowing.
Ask it what it remembers, and distrust the answer. These systems produce fluent, plausible descriptions of their own workings that are not reports of their own workings. Treat the reply as marketing copy generated on demand, not as documentation.
Check what a fresh start actually resets. Look for the difference between starting a new conversation, clearing memory, and deleting the account. On some apps the first two are the same action; on others clearing memory leaves the stored history untouched.
Memory is a retention decision before it is a feature
An app that remembers you between sessions is an app that stores something about you between sessions. That is not a criticism — it is definitionally what the feature is — but it means every memory claim is also a claim about a database on somebody’s server, with a lifespan set by policy rather than by your use of the app.
Two questions follow, and neither is answered on a feature page. The first is how long that stored material lives once you stop using the app, which is the subject of what a companion app operator actually holds. The second is what the stored material may be used for besides talking to you, which is a policy question covered in how to read a companion app’s privacy policy.
The practical version: the more an app remembers, the more there is to think about later, and the two properties are not separable by choosing carefully.
What a memory claim never promises
It does not promise a transcript you can read. Remembering and archiving are different products. An app can carry facts between conversations while offering you no way to retrieve, search, or export what was said.
It does not promise durability. Memory behaviour is one of the things a release can change without asking you, which is why it is worth re-running the fact-planting test after a major update. The general pattern is described in what an app update can change without asking.
It does not promise accuracy. An app that remembers wrongly is behaving exactly as designed, and confident wrongness is the normal failure mode rather than a fault to report.
It does not promise continuity survives anything else. Changing accounts, changing devices, or reinstalling can each reset it. What breaks and why is the subject of when a companion app stops remembering.
Where the testing runs out
None of this reveals the mechanism, and it is not meant to. You can establish that a fact survived a week and a new conversation; you cannot establish how, where, or in what form it was kept, and any confident story about that is a guess dressed as a finding.
What the tests do give you is the only thing a marketing page cannot: a behaviour you have observed rather than one you were promised. If you plant a fact, wait, and ask — and the answer comes back right, twice — you know what you are dealing with. If it comes back wrong, or comes back as a fluent paragraph about how much it treasures your conversations, you also know.