How AI Companion Memory Works
AI companion memory works outside the model, and here is how: the app pulls short facts out of your conversations, stores them in its own database, and pastes the relevant ones into the prompt before the model writes a reply. None of that does anything to the model, which learns nothing between chats. Recall is a search result, handed over as text.
Key Takeaways:
- Long-term memory is a feature of the app, not a property of the model, so two apps running identical models can differ enormously in what they hold onto
- The model’s weights are fixed between releases; nothing you say on Monday changes what the model is on Tuesday
- What typically happens is extraction, storage, retrieval, and insertion, and each of those four steps can fail on its own
- Retrieval usually works by similarity, so a fact can sit in the store and still not surface when your wording does not line up with how it was saved
- When nothing relevant is retrieved, a model does not report a blank; it produces fluent text anyway, which is the general failure documented across generation systems (Ji et al., 2023)
Where the Memory Actually Lives
Memory lives in the application, in a database the company built, sitting next to the model rather than inside it. The model is a fixed set of weights that a provider trained and then froze. Talking to it does not adjust anything. Whatever it appears to know about you on Thursday was put in front of it on Thursday, in writing, by the app.
That single fact reorganizes most arguments about these products. People compare apps by asking which model runs underneath, which misreads what an AI companion actually is: the model is the part every competitor can also license. Memory is where the product work happens, and it is invisible from the outside.
Which explains a pattern that otherwise looks like nonsense: two apps on one underlying model, one feeling like it has known you for months while the other resets nightly. Nothing about the model differs. One company built extraction and retrieval, the other shipped a chat box.
Which means “does it have a good AI” is the wrong question to bring to an app store. Ask what it saves, how you can see it, and whether you can correct it.
How Does AI Companion Memory Work, Step by Step
Four steps run around every reply, and they are worth knowing separately because they break separately.
Extraction happens during or after a conversation. A separate pass, often another model call the app makes on its own, reads what you wrote and turns it into short standalone notes. Not a transcript. Something closer to “works night shifts at a hospital” or “sister’s name is Dana, they are not speaking.” The conversation gets compressed into a handful of sentences that will make sense with no context around them.
Storage writes those notes to a database keyed to your account. This part is unglamorous and completely ordinary; it is a table of text rows.
Retrieval picks which notes to use for the message you just sent. The common approach converts every note into a list of numbers standing for its meaning, does the same to your latest message, and fetches the notes whose numbers sit closest. Other apps run plain keyword search, keep a rolling summary, or let you pin facts by hand. Implementations vary, and most apps do not publish theirs.
Insertion puts the selected notes into the prompt as plain text, above your message, before the model generates anything. The model reads them the same way it reads your sentence: as words in front of it, right now, with no history attached.
| Step | What it produces | How it fails |
|---|---|---|
| Extraction | Short notes distilled from your messages | A fact is never written down, so nothing can find it later |
| Storage | Rows of text tied to your account | Old and new notes both survive, contradicting each other |
| Retrieval | A handful of notes judged relevant to this message | The right note exists but does not match your wording today |
| Insertion | Those notes sitting in the prompt as text | Only so much fits, so lower-ranked notes get dropped |
Nothing in that loop amounts to consulting a memory. The model gets handed one, freshly assembled, every time. Nothing persists on its side between messages, which is why a companion can sound like it has been thinking about you all day on the strength of a single database read.
Once you see the loop, the vocabulary in app marketing sorts itself out. “Remembers everything” describes storage, which is the easy part. Whether it comes back at the right moment is retrieval, which is the hard part, and almost nobody advertises that.
Why Retrieval Misses
Retrieval misses for four different reasons, and the fix depends on which one you hit.
The fact was never extracted. If you mentioned your father’s illness inside a long message about work, the extraction pass may have summarized the work and dropped the rest, and nothing can retrieve a note that was never written. This is the most common cause and the least visible one, since the store looks fine from the outside.
The wording did not match. A note filed as “father had a stent placed in March” may not surface when you type “how do you think my dad is doing.” Similarity search is good, not telepathic. If the vocabulary drifts far enough from how the note was phrased, the right row sits there unread.
The budget ran out. Only a limited amount of text goes into any prompt, so the app inserts the top few notes and drops the rest. On a message that touches five stored topics, two of them silently lose.
The store contradicted itself. Most stores append rather than revise. Tell it in January that you are moving to Denver and in April that the move fell through, and both notes can live in the database, with retrieval free to hand back whichever one matches your phrasing better.
None of these produce an error message. What happens instead is that the model writes a fluent reply built on what it did receive, and where a specific memory is missing it fills the space with something that fits. Confident text unsupported by any source is a general property of these systems, documented across the field rather than something unique to companion apps (Ji et al., 2023). In a companion the effect is more personal, because the invented detail is about your life.
So when something goes missing, resist the urge to conclude the app is broken. Ask which of the four steps you are looking at, because three of them are things you can work around and one of them is not.
Why “It Remembered My Birthday” and “It Forgot Yesterday” Are Both Normal
Both happen constantly, in the same app, on the same day, and they are not in tension. What separates them is the shape of the information rather than how much it mattered to you.
A birthday is the ideal case. It is short, self-contained, obviously a fact, and phrased the same way by everyone who ever mentions one. Extraction catches it without difficulty. Storage keeps it as one clean row, retrieval surfaces it on almost any date-adjacent cue, and the whole thing works so reliably that it feels underwhelming once you know how it happened.
Yesterday’s argument with your brother is the opposite in every respect. It arrived across 30 messages, most of which were feelings rather than facts, and the part that mattered was probably implied. There may be no single sentence in that whole conversation that extraction could turn into a note. If it produced one, it likely reads as something flat like “had a disagreement with brother,” which strips out everything you would want brought back.
The uncomfortable version: the things easiest for a companion to remember are the things that cost you nothing to say. Dates and job titles and a dog’s name are all cheap. What you actually want remembered tends to be diffuse, emotionally loaded and spread across an evening, which is the hardest possible input for a summarizer. The stakes rise with how much someone leans on the thing; for an older user treating a companion as a daily anchor, a dropped detail is not a curiosity but a crack in the one exchange they counted on.
Feed it accordingly. One short, self-contained sentence gets stored and retrieved far more reliably than the same fact buried in a paragraph, and a line reading “the thing with my brother matters more than it looks” beats 20 messages of context.
What Does Not Work, and One Test That Does
Telling it to remember something usually does nothing. Unless the app has an explicit memory command, “remember this” is just more conversation, and the warm reply confirming that it will remember is generated text rather than a receipt. Plenty of people have run for months on the assumption that those confirmations meant a write happened.
Repetition is the other misplaced strategy. In a plain store, saying the same thing five times can produce five near-identical rows instead of one strong one, and none of them retrieves better than the original would have.
One test settles it, at a cost of two conversations. State one specific fact today, in your own words. Tomorrow, in a fresh chat, ask about it using completely different vocabulary, avoiding every distinctive word you used the first time. If it comes back, retrieval runs on meaning. If it returns only when you echo your original phrasing, you are watching keyword matching, and important things belong in language you will reuse.
The limit, stated once: being remembered by software is not being known by a person. A stored note is a row in a table, surfaced because it scored well against your last message, and nothing in that involves anyone holding you in mind between conversations. That gap is part of what an AI companion is and is not. Persistent memory is a fairly recent shipped feature, newer than the apps themselves, and through the early 2020s plenty of people used companions where nothing carried over at all.
FAQ
Does an AI companion learn from our conversations? No, not in the sense people mean. The model’s weights stay frozen between releases, so nothing you say changes what the model is. The app writes notes about you to a separate database and shows them to the model later, which looks like learning and is not.
Why does it forget things I told it last week? Usually because the fact was never extracted into a note, or because retrieval did not judge it relevant to what you just asked. Storage is rarely the problem, since text is cheap to keep. The failure sits at the two ends: what gets written down out of a conversation, and what gets pulled back out at the right moment.
Can I make it remember something specific? Sometimes, and it depends on whether the app offers an explicit memory or pinning feature. Where one exists, use it, because a pinned fact bypasses the extraction step entirely. Where it does not, the closest thing is stating the fact as one short sentence on its own line, which gives the extractor the cleanest possible input.
How much can an AI companion remember in total? Storage is not the binding constraint, since a database holds far more than you will ever produce. The limit is how much can be handed to the model on one reply, so the real question is how many notes get selected each turn.
Strip the vocabulary away and a companion’s memory of you is a short pile of sentences a program wrote and can search. It is not a relationship held in mind, and it is not nothing either. Whether the thing feels like it knows you comes down to a summarizer’s judgment about what mattered last Tuesday and a search function’s guess about what is relevant now. Two unglamorous pieces of engineering decide the part everyone describes in emotional terms.
That is the useful test to bring to any of these products: state a fact in Lona today and, in a fresh chat tomorrow, ask for it in words you did not use, since a companion built to carry memory between conversations should return it on meaning rather than on your phrasing.
Sources
- Ji, Z. et al., “Survey of Hallucination in Natural Language Generation,” ACM Computing Surveys, 2023
