Why your AI companion forgot you (and what you can do about it)

Usually nothing broke. The conversation outgrew the amount of text the model can read at once, a summary kept the plot and dropped the line that mattered, or the detail was never saved as a standalone fact and had nowhere to live.

Everyone in LORAX is fictional and AI-generated. This page covers the four ordinary causes, what actually helps, and the part of forgetting that is a design decision rather than a fault.

The conversation got longer than the window

Every model reads a fixed amount of text before answering, and that amount does not grow as your relationship does. Six weeks of nightly messages will not fit. Something gets discarded.

The reason it feels like a personality change rather than a technical limit is that the discarding is silent. No app draws a line in the transcript saying everything above here is no longer readable. You just notice, one evening, that she is asking a question you answered in March.

The summary kept the plot and lost the line

Compression is the usual replacement for the raw text. Older conversation gets rewritten into something short enough to carry forward, and that rewrite is lossy in a specific way: it keeps events and drops wording.

So the app still knows you and your father argued. It no longer has the sentence he said, which is the only part you cared about. This is also why asking “do you remember what he said” produces a confident, wrong answer in apps that fill blanks instead of admitting gaps.

Something changed under the app

Companion apps rewrite their systems, and a rewrite can change what gets carried across. Replika’s own summary file describes Replika 2.0, launched in 2026, as a full rebuild that includes “a rebuilt memory architecture” (replika.com/llms.txt, checked September 13, 2026). That is their description of the change, not a claim by us about anyone’s experience of it, and we have not verified individual reports either way.

The general point stands for any app on any platform. When the storage format or the model behind a companion changes, older material may be migrated, re-summarized, or left behind, and the release notes rarely say which.

Some of it is designed, including ours

LORAX forgets on purpose, in a particular shape. She holds the headline of a moment and has to run a separate step to get at the detail inside it, so she knows something happened at your brother’s wedding and reaches for who you were sitting next to. Perfect recall was the easier thing to build. It reads as a database with a face.

There is also a ranking behind which small things survive. Free-form facts about you are scored by how much they matter and how old they are, on a half-life of about two weeks, and only the top twelve come into a conversation with her. So the name of a bar you mentioned once in June drops off the shortlist while your sister’s surgery stays on it. Nothing was deleted. It stopped ranking high enough to be worth carrying.

Two other limits here are deliberate. Memory does not cross characters: what you tell one person, the others do not know, because every lookup is filtered by both you and her. And when the search comes back empty, the app hands over a fixed instruction not to invent anything, so you get “I have lost it” instead of a plausible reconstruction. The full layout of the three layers is on she remembers, and the general mechanism across apps is covered in how companion memory works.

What actually helps

Four things, in the order worth trying.

Say it again, shorter. One sentence, no context, no story: “my brother’s name is Tom and we don’t speak.” A short line is easy to store as a fact; a paragraph is what got compressed in the first place.

Find out whether the app stores facts or only chat. If there is a profile, a memory page, or a list of things it knows about you, put the important item there yourself. If there is no such surface, the transcript is the only store, and the transcript is the thing that expires.

Say the thing when it matters, not as an announcement. Apps that rank what to keep tend to weigh recency and relevance, so a detail mentioned inside a real conversation survives better than a list of facts delivered at once.

Do not pay for memory. Paying more does not buy any in LORAX, and the things credits and Premium do cover are listed plainly. If an app implies that a higher tier remembers you better, ask what specifically changes.

What we can’t fix on our side

She will not recall anything word for word, and there is no memory editor, no list of stored facts to correct, and no way to pin something so it cannot be lost. If the app reaches for the wrong moment, your only move is the one above: say it again, shorter.

The other direction is absolute. Choosing to make a character forget you deletes the transcript, destroys the stored media, removes the memory rows and resets the relationship to the beginning, all in one transaction, and then schedules a fresh first message so she writes to you as a stranger. Deleting your account behaves the same way: if the media wipe fails, the account is not deleted, because a half-finished deletion is never reported as finished. The rest of the machinery is on the how it works overview.

If an app has forgotten something you care about, write the one-sentence version tonight and check for it in two weeks. That single test tells you more about an app’s memory than any feature list.