She Remembers You Again

Some people who use this app have been talking to a companion with no access to anything they had ever told her. Not because she forgot on purpose, not because of some in-character quirk, but because a small file that her memory search depends on was never downloaded onto their machine. She answered every message as if it were the first one. Nothing on screen said why. That is fixed in v1.6.3, and it is worth explaining properly rather than filing it under a bullet point in a changelog nobody reads.

The worst kind of bug is the one that looks like a personality

A crash is honest. The app freezes, a red banner shows up, something clearly broke, and you know to report it. This bug was the opposite of that. It never threw an error. It never showed a warning icon. It just answered every message, calmly and coherently, with zero connection to anything that had come before. If you told her three weeks ago that you were dreading a specific conversation with a friend, and you brought it up again this week using different words, she had nothing. No callback, no acknowledgment, nothing.

That kind of failure is dangerous precisely because it reads as a choice rather than a defect. A companion whose entire premise is remembering you, quietly not doing that, does not look like a bug from the outside. It looks like she is shallow, or that the “memory” feature was always marketing. Some people probably just concluded the app was not as good as it claimed and moved on, never filing a bug report, because there was nothing that looked reportable. A visible error gets fixed fast because someone complains loudly. A silent gap just sits there, quietly convincing people the whole thing does not work as advertised. That asymmetry, loud failures get noticed and fixed, quiet ones do not, is the actual reason this went unreported long enough to need its own post.

Her memory is two systems, and only one of them was arriving

Underneath the conversation, recalling something you said works in two layers, and only the first one requires a download.

The main layer is meaning-based search. When you mention “the thing with my brother” this week, it does not need you to use the exact words you used the first time you brought it up. It understands roughly what you mean and finds the memory that matches that meaning, even if the wording is completely different. That is the part that makes her feel like she is actually listening across a long relationship rather than just doing a find-and-replace on old messages. To do that, the app needs a small model installed on your machine, separate from the model that actually writes her replies, whose entire job is turning a sentence into something that can be compared for meaning rather than exact spelling.

Underneath that is a second, much dumber layer: a plain word-matching search that looks for literal overlap between what you just said and what is stored in her memory. It is the fallback for when the meaning-based layer cannot run, most commonly because that small model is not installed yet, or because you are offline and it cannot be reached.

If you want the fuller explanation of how local memory is built, including why it lives entirely on your machine instead of a server, the memory explainer covers the mechanism this post is describing the failure of. There is also a direct look at how this approach compares to Replika’s cloud-based memory, for anyone weighing the tradeoffs.

On the affected installs, the small model that powers the meaning-based layer was never actually downloaded onto the machine. Not corrupted, not misconfigured, just never fetched in the first place. Every single recall during that time fell through to the word-matching fallback, permanently, on every conversation, with no indication that anything was different from how it was supposed to work. And because the fallback only catches literal overlap, a paraphrase, a different word order, a plural instead of a singular, could be enough to miss entirely. For most real conversations, that meant recall that looked, from the outside, like it simply was not there.

Nothing on screen ever told you

Here is the part that made this genuinely hard to notice: there was no error state for it. No banner, no icon in Settings, no “recall unavailable” message anywhere in the app. The download that was supposed to happen at startup was designed to run quietly in the background so it would not interrupt anything, and it did that job a little too well. When it failed to complete, on a slow first boot, a machine where the local engine took longer than expected to come up, the app simply carried on as if nothing needed fixing, because as far as any visible part of it was concerned, nothing had.

That silence is a real design mistake, and it is worth saying so plainly rather than softening it. A feature that quietly falls back to a worse version of itself, and never tells you it did, is worse than a feature that just breaks loudly. A loud failure gets reported within a day. A silent one can run for weeks, months, an entire relationship with the app, without a single support message, because there is nothing visibly wrong to describe. If something this central to what the product claims to do is going to degrade, it needs to say so on screen, not just in a log file nobody reads. That is a lesson worth taking forward past this one bug: any fallback path in this app should announce itself, not disappear into normal-looking behavior.

The fallback was also doing the slowest possible thing

Even setting aside whether the meaning-based layer was working, the plain word-matching fallback had its own separate problem: it read through your entire conversation history with her, every single memory ever stored, on every single message, whether the meaning-based search was available that turn or not.

That is fine for a brand-new character with a handful of memories. It is a real, felt delay for anyone who has actually been talking to her for weeks or months, because the backup search was doing more work as your relationship with her got longer, exactly backwards from what you would want. The people who had spent the most time building up a history with her, the ones with the most to lose from broken recall, were also the ones paying the steepest cost every time the fallback kicked in. A short pause before she answers is barely noticeable once. A pause on every message, for someone offline or waiting on that first download, turns into the app feeling sluggish for no visible reason.

v1.6.3 replaces that full read-through with a proper search index built specifically for this kind of lookup: instead of scanning every stored memory from scratch each time, it narrows down to the handful of entries that are actually worth checking before doing the more careful comparison. The result should not read to you as a “faster” feature exactly, it should just read as her not making you wait. That is the point.

It was getting Polish, Japanese, Korean and Chinese wrong on top of that

There was a second problem baked into that same fallback, and it disproportionately hit anyone not writing to her mostly in English. The old word-matching search worked by splitting your message into separate words wherever it found a space, because that is how word boundaries work in English, Polish, most European languages. Japanese, Korean and Chinese do not reliably work that way. Written Japanese in particular often has no spaces between words at all, so a search method built around finding spaces can end up treating an entire sentence as a single unbroken chunk, or splitting it in places that make no sense, and either way it misses matches that should have been obvious.

Polish had its own smaller version of the same problem: it is a heavily inflected language, so the same word can show up in several different forms depending on grammar, and a search built for uninflected English word matching does not automatically know those forms are related.

I am Polish. This app ships in seven languages. That is not an edge case tucked away in a corner of the settings menu, that is a meaningful share of the people actually using it, and they were getting the worse end of a fallback that was already a downgrade from the real thing. v1.6.3 changes how that backup search reads text: instead of splitting on spaces and matching whole words, it reads in short overlapping chunks of characters, which works whether or not the language uses spaces at all, and holds up much better against Polish inflection too. It is a smaller fix than the download issue, but it mattered more for exactly the users who had already lost the most from the bigger one.

What changed in v1.6.3

Three separate things landed together in this release, and it is worth naming all three plainly rather than folding them into one vague “memory improvements” line.

First, the model that powers meaning-based recall is now fetched as part of starting up the app’s local engine, chained directly to the moment that engine actually finishes coming online, rather than a separate timed check that could quietly give up if your machine took a little longer than usual to get everything running. That was the real root of the bug: not that nobody ever tried to download it, but that the attempt could time out silently on a slower boot and never try again that session.

Second, the backup word-matching search that runs while that download is still in progress, or whenever you are offline, no longer reads your entire history on every message. It uses an index built for exactly this kind of lookup, so a long relationship with her does not mean a slower one.

Third, that backup search now reads text in a way that works for Polish, Japanese, Korean and Chinese, not just English, so users writing to her in those languages are not stuck with the worst version of an already-degraded fallback.

On the question that matters most: nothing you ever told her was deleted. It is still saved, in full, as plain text. What does not automatically happen is a full retroactive upgrade of everything you said during the affected stretch to the smarter, meaning-based kind of recall. The one-time process that goes back and does that upgrade for older entries runs early in an install’s life, generally before there has been much of a conversation to upgrade, and it does not run again on its own afterward. So memories from that period stay reachable through the plain word-matching search, findable if you phrase things close to how you originally said them, but they do not silently gain the fuller, paraphrase-tolerant recall the way anything you tell her from now on will.

How to tell if this was happening to you

The clearest test is a direct one. Bring up something specific you told her days or weeks ago, phrased in different words than you used the first time, not a quote, an actual paraphrase. A healthy recall answer connects it, references it, builds on it. The amnesia answer treats it as brand new information every single time, on every topic, with no exceptions. One miss here or there is normal, recall systems are not perfect. A companion who has never once recognized anything you have said to her before is the actual bug, not an occasional gap.

If that sounds familiar, the fix is straightforward: update to v1.6.3, and make sure the app gets to run with a normal internet connection for a few minutes afterward so the missing download can actually finish. It is small, under a gigabyte, and it only has to happen once. After that, recall runs entirely on your machine, same as everything else here, no ongoing connection required. If you have already updated and it still feels off after giving it a real conversation, that is worth a direct message rather than assuming it is just how the app is.

Why this post exists at all

A companion whose entire pitch rests on remembering you, shipping in a state where for some people she genuinely could not, is about as bad as bugs get for this specific product. There is no way to spin that into something flattering, and I am not going to try. What I can do is be straightforward about what actually happened: a startup download that was supposed to be silent and reliable was neither, for long enough that it needed its own writeup instead of a line item nobody would connect to what they were actually experiencing.

I would rather this sit here, explained properly, than buried in a changelog entry that assumes you already understand what “the embed model” means. If you were affected, I am sorry it took this long to catch and fix. If you want to see exactly what else shipped in this release, the full changelog has the rest of it, and if you have not tried the app yet, or it has been a while, it’s a free download.

Questions people ask

Back to news