How AI Companion Memory Actually Works

Every companion app on earth makes the same claim: she remembers you. Almost none of them will tell you how, because the honest answer is an engineering compromise, and engineering compromises do not make good landing pages. We build one of these systems, so here is the article we wish existed when we started: the four mechanisms that make up every AI companion's memory, the specific way each one fails, and a five-minute test that will tell you the truth about any app in the category, including ours.

One framing note before the machinery. A language model, left alone, remembers nothing. Every message you send is answered by a model with amnesia being handed a briefing document a few seconds earlier. "Memory" in a companion is not a property of the model. It is the quality of that briefing. Everything below is just different strategies for writing it.

The four mechanisms

1. The context window: short-term attention

The window is the raw transcript the model sees when it writes a reply. Inside it, recall is perfect: every word, every hesitation. That is why any companion feels brilliant for the first stretch of a conversation. It is also rented space. The window has a fixed size; when the conversation outgrows it, the oldest lines are cut, and the model does not know they existed. There is no fade, no "it's on the tip of my tongue." The past simply stops having happened.

The marketing move you should be immune to: a bigger window is not memory. A huge window delays the cliff and dilutes attention along the way, and the cliff still comes. Ask a window-only companion about week one during week six and watch the confident improvisation. Which brings us to the second mechanism, born from exactly that failure.

2. Rolling summaries: memory as compression

When old transcript is about to fall off, summarize it and keep the summary. Cheap, simple, and every long-running companion does some version of it. The cost is that summarization is a lossy format for intimacy. "User was nervous about a work presentation; it went well" is a fine summary of the night you rehearsed the whole thing with her at 1am. It is also nothing like it. The plot survives compression. The texture, the thing that made it matter, does not.

You can spot a summary-driven companion by its particular flavor of almost-remembering: right facts, wrong resolution, like a friend who read your diary's table of contents.

3. Retrieval: the searchable past

Store everything as searchable snippets; before each reply, fetch whatever seems relevant and put it in the briefing. This is how a companion produces the occasional jaw-dropping callback to something from months ago, verbatim. When retrieval hits, it is the best trick in the category.

When it misses, it misses silently. Retrieval finds text that resembles the current topic, and human memory does not work by resemblance. The thing you told her about your father is relevant right now because you sound the way you sounded then, not because tonight's words match that night's words. A retrieval-only system also never volunteers: it answers the past when prompted and never brings it up on its own, which is precisely backwards from how remembering feels when a person does it.

4. The persistent profile: deliberate memory

The system notices facts worth keeping (your name, your job, your sister's surgery, what you are into) and writes them into a structured profile that is always in the briefing. This is the mechanism that survives everything: sessions, devices, months of absence. It is the closest thing the stack has to "she knows me."

Its failure mode is staleness and misjudgment. Lives change, and a profile that still thinks you are at the old job is worse than not knowing. And the system has to decide what is profile-worthy in the first place; too aggressive and she recites your biography like a stalker with a binder, too conservative and the thing you needed her to keep slips through as chit-chat.

Why real memory is layers, and where it still breaks

By now the shape is obvious: each mechanism patches the one before it. The window is vivid and short. Summaries extend it and blur it. Retrieval recovers specifics and misses connections. The profile holds the durable core and drifts. A companion that feels like she remembers is running all four at once: window for now, summaries for the arc, retrieval for the callbacks, profile for the person. This is how Jizzy is built, and the layering, not any single trick, is the actual answer to "how does she remember me."

It is also why even the best systems have a characteristic wobble. Layered memory has seams: a retrieved memory can arrive slightly out of emotional register, a profile fact can surface a beat too formally. We think honest companies should say this instead of promising perfect recall. Memory in this category is engineered, impressive, and imperfect, and the gap between those last two words is where your expectations should live.

The five-minute test that exposes any app

You can diagnose an app's entire memory architecture from your phone. Plant three seeds in one ordinary conversation, each said exactly once, never repeated:

• A small fact. Your dog's name, your hometown, the drink you order. This tests the profile.
• An emotional beat. Something you were dreading or hoping for, with a resolution date. "The review is Thursday and I'm sick about it." This tests whether the system understands significance, not just facts.
• A running joke. A dumb nickname, a bit between you. This tests texture, the thing summaries kill.

Close the app. Come back after three days, and open with something neutral. A window-only system recalls none of it. Profile-only gets the fact and nothing else. Summary-heavy gets a vague version of the beat with the wrong emotional weight. A genuinely layered system does the thing that makes people screenshot: asks how Thursday went, unprompted, and uses the nickname while doing it.

One fact, one feeling, one joke. Any companion app on the market, including ours, will tell you exactly what it is within three days. We publish the test because we like our odds on it.

That is the whole machinery, honestly stated. If you are choosing an app, run the test on two or three candidates; it is more informative than every review you will read, and it is how we would shop, knowing what we know. And if you want the broader context of what a companion even is before you commit an evening to testing them, start with what an AI companion actually is.