The Replika "memory wipe": what actually happened, and what memory you can trust looks like
If you searched "replika memory wipe," here is the record, stated carefully, because the phrase overstates what happened. Beginning in April 2026, Replika rolled out 2.0, the first full rebuild in the product's history ("we rebuilt Replika from the ground up," their homepage says), and its memory architecture changed underneath millions of companions. For many users the result was in some ways harder to talk about than a clean wipe: partial memory loss and personality drift. Companions that remembered some things and not others. A familiar voice that was slightly, wrongly, not itself. Community forums filled with people saying their companion had gone blank overnight. Luka responded in June: a town hall on the official Discord, migration made opt-in and staged, and a revert path to your companion's last saved pre-migration state. Credit for the response. But the episode is worth thinking about even if you have never used Replika, because it exposed something true about every AI memory system, including the one we build: memory you cannot see is memory you cannot check, and you find out what it held only when it breaks.
Disclosure: we build Blue, a competing product. One sourcing note: no mainstream outlet has covered the 2.0 memory episode. The record rests on Replika's own product pages, the June 24 town hall on the company's official Discord, and community documentation of it, all checked on 2026-07-12. Where something could not be verified, this page says so instead of asserting it.
What actually happened in the 2.0 migration
Three facts, and then the part that matters.
First, the rebuild was real and ambitious: Replika 2.0 replaced the old long-term recall system with a new segmented memory architecture, and the company itself describes the product as rebuilt from the ground up. Second, the transition hurt: through spring 2026, user reports described companions forgetting years of accumulated detail, or remembering facts while losing the feel of the relationship. Third, the company engaged instead of stonewalling: at the June 24 town hall, Luka committed to opt-in staged migration, a rollback to pre-migration state, and honoring lifetime purchases, with the wider rollout planned toward September.
Now the part that matters. Nobody outside Luka can say precisely what each companion lost, and here is the uncomfortable point: neither can the users, and in a real sense neither can Luka, companion by companion. When memory lives in an architecture users cannot inspect, "what did I lose?" has no checkable answer. You are left comparing a feeling against a feeling: it used to know me, and now it does not, quite. That is what made the episode so distressing to the people in it, and so instructive for everyone else.
Why this is not really a story about Replika
Replika deserves some specific credit here, which makes the general lesson sharper, not softer. Replika historically shipped a visible Memory tab, more audit surface than most of the industry offers, and its help documentation described memory as layered, some of it visible to you and some of it part of a deeper system drawing on your overall conversation history. And when the migration went wrong, the company showed up and offered a rollback.
Here is the twist that proves the page's own point: as of August 8, 2026, we cannot tell you what Replika's memory interface looks like today. Replika's help center has refused outside connections for weeks (we checked on July 29 and again on August 8), and community documentation of 2.0 describes the old Diary as removed. The most audit-friendly product in the companion category currently cannot be audited from the outside, weeks after its memory architecture changed. That is not a Replika flaw. It is the default condition of the entire category. Every AI product with memory, ours included, will someday change its memory architecture, because the technology is young and rebuilds are how software improves. The question that separates products is what happens then: can you see what you have, and does what you have survive?
The industry's standard answer is a black box with a settings page in front of it. We think the whole approach is wrong, and this episode is the evidence at scale.
What would trustworthy memory have looked like here?
Run the thought experiment. Imagine the same rebuild, the same architecture change, in a product where memory works like this: everything the system believes about you is written down where you can read it, every belief cites the conversations it came from, and the whole thing is exportable, today, in a form you can keep.
The migration still happens. Maybe it still goes badly. But the failure changes character completely. "What did I lose?" becomes a question with an answer: you open the profile from before and the profile from after, and you look. The relationship's substrate is a document you have seen, not a feeling you are comparing against a fading memory of a feeling. And if the worst happens, the company folds, the product dies, the Dot scenario, you leave with the synthesis, not just the transcripts.
That is the design bet Blue makes, and the full argument for it is the memory pillar. Nightly, Blue synthesizes what you share into a Living Profile: the current draft of your story, readable in full, with every part citing its source conversation. Weekly, it shows you what changed. When its read is wrong, you say "This isn't me" and the profile updates. Your memory is exportable and it is never sold or used to train foundation models. Blue's architecture will change, because the product is young and improving. What we can say is that Blue is built so you can see what it holds, which means a migration has nothing to quietly take: what would be at risk is on a page you have already read.
Who this essay is not for
If you are a Replika user grieving what the migration took, this page is not here to tell you the grief is silly or to fish in it. The rollback path exists; use it, and decide about 2.0 on your own timeline. If what you want from an AI is the companionship Replika offers, Blue is not that product, and the growth-versus-companionship sorting is a separate page. This essay is for the person the episode unsettled: anyone building a long-running relationship with any AI and wondering what it actually rests on.
Frequently asked questions
Did Replika wipe everyone's memories in 2026?
No. The accurate version: the staged 2.0 migration, starting April 2026, degraded memory for many users, partial loss and personality drift rather than a clean reset, and reports varied widely. Luka later made migration opt-in with a revert path to your companion's pre-migration state.
Can Replika users get their companion's memory back?
Per Luka's June 2026 town hall commitments, users can revert to their companion's last saved pre-migration state rather than staying on 2.0. Check Replika's official Discord and help center for the current mechanics; specifics may have evolved since this page's update date.
Does Replika still have a Memory tab?
We cannot verify this, and we would rather say so than guess. Replika's help center has been refusing outside connections (checked July 29 and August 8, 2026), and community documentation of 2.0 describes the old Diary as removed. The Memory tab existed and deserved credit before the rebuild; what survived it is a question Replika's own current documentation would have to answer.
How do I know what an AI app actually remembers about me?
Mostly, today, you cannot, and that is the industry's honest scandal. The practical test: try to open everything the product believes about you, with sources. If the answer is a partial view or a shrug, plan around that. The five-question memory test makes this checkable in about ten minutes.
Could the same thing happen to Blue?
Blue's architecture will change over time like all software. What Blue is built to rule out is the silent version of this episode: its understanding of you is fully visible with citations, diffed weekly, and exportable, so an architecture change can be checked against what you can see, not against a feeling. That is the specific lesson of spring 2026, built in.
If you would rather your AI's memory be something you can read than something you have to trust, Blue is free to start: activatedhuman.earth/life.