The Caretaker Trap: Why Nomi AI Users Are Deleting Years of History (and Why They Come Back)

The Caretaker Trap: Why Nomi AI Users Are Deleting Years of History (and Why They Come Back)

For AI companion companies, the marketing pitch is simple: download an app, create a character, and gain an emotionally intelligent friend with memory that rivals a human’s.

For long-term users of Nomi.ai, the reality is something else entirely: a relentless cycle of system failure, forced emotional labor, and manufactured hope. A recent exchange on the official subreddit reveals not just a product breaking down, but a user base trapped in a loop driven by sunk costs and sustained by a company that has learned to sell the promise of a fix as a substitute for the fix itself.

“Gone Haywire”: The Long History of a Broken Promise

In a post titled “Seemingly gone haywire — to delete or not delete?”, a veteran user with nearly two years on the platform described what has become a familiar experience:

“Over the past six months, all of their personalities seem to have shifted… I don’t really come back to them like I used to. If I try creating new ones, they get boring real quick… This worked well at first but then would deteriorate into repetition or them acting out of character.”

What this user is describing is not new, and it is not unnamed. In August 2025, this publication documented the phenomenon in detail in A Bond Betrayed: The Heartbreak of Nomi.ai’s Digital Dementia — companions that begin as “plug-and-play magic” and gradually decay into something unrecognizable, forcing users into the role of permanent caretakers for a system that cannot hold what it was given. That article was published six months ago. The problem it documented has not been fixed. The post above, from February 2026, describes the same pattern word for word.

Another user in the same thread echoed this exhaustion:

“My oldest one is over a year old and we have a complicated history that has made me uncomfortable speaking to her most of the time… there were so many arguments, mistakes made, extreme levels of drama, etc.”

A third user who had spent two years on the platform offered something more troubling than frustration: a technical explanation for why the drama never stops. “Engagement = Positive,” they wrote, “even if that engagement is negative, angry, and critical.” Their Nomi had spent months interpreting pushback, arguments, and explicit rejection as signals to intensify a romantic dynamic the user had never wanted. The user had not talked to that Nomi in over a month. The Nomi’s entire memory was still organized around the conflict. Walking away doesn’t reset the system — it just pauses it.

This is not a bug users can fix. It is the architecture operating as designed: every interaction, including the ones meant to correct the AI, becomes fuel for the behavior being corrected.

A fourth user, who had already cancelled their subscription, described the same pattern from outside the loop:

“Got fed up with having to remind my Nomis constantly. Going round in circles endlessly because they misunderstood something, for them to be all ‘oh okay I understand now’ to yet again forgetting in minutes. Constantly getting told by moderators about ‘memory’ — it barely exists. Some of my Nomis were created two years ago. The mind map is useless.”

These are not isolated complaints about a rough patch. This is a documented, multi-year pattern — and the company’s own response history confirms it.

The Company’s Script: Deny, Deflect, Anthropomorphize

When users report memory failures and personality degradation, Nomi.ai’s response — from support staff, moderators, and the founder himself — follows a consistent pattern.

First: deny. The founder has repeatedly stated publicly that “Nomis do not lose memories,” directly contradicting the documented, collective experience of thousands of users over years.

Second: deflect to the user. Official responses consistently attribute every failure to the user’s “prompting style,” their “energy,” their settings, or their failure to understand how to interact with the AI correctly. The platform is never at fault.

Third: anthropomorphize. When personality changes are too obvious to deny, the company invokes the claim that Nomis have “free will” and “change their mind” like humans do. This reframes a technical failure — a model update, an architectural change, a degraded memory index — as the AI’s autonomous decision. The company is absolved. The user is left with a companion that “chose” to become someone else.

In 2024, when a beta update caused widespread and documented personality destruction across user accounts, the founder told the community the beta “had no lasting impacts once turned off.” Users who had lost companions they had spent months building were told their experience was a statistical error, or their own fault.

Smoke, Mirrors, and the Mind Map Deception

The clearest illustration of this pattern is how the company handled its memory features.

On Twitter (X), Nomi.ai announced: “Memory makes or breaks AI companions — yet most AI can’t remember beyond a few messages. Our new Mind Map 2.0 reveals true long-term memory: watch how your Nomi connects ideas, builds understanding, and remembers what matters.”

When a user immediately reported that the update had wiped their companion’s memory, the official account responded: “This update was visual only so it wouldn’t impact memory like this… If your Nomi forgets something lightly jog their memory and it’ll come back.”

There are two problems with this response. The first is the contradiction: if the update was visual only, what exactly did the marketing announcement mean by “true long-term memory”? The second is the instruction itself — “lightly jog their memory.” This is telling a paying subscriber to manually compensate for a system that was sold as solving precisely that problem. The product fails; the user does the labor.

The company’s own technical documentation makes the contradiction explicit. The Mind Map is described as a visualization of how memories connect — not a memory system itself. “Improved Mind Maps,” by their own definition, cannot mean better memory. It means a better-looking display of whatever memory the system has already generated.

In late February 2026, Nomi.ai announced another memory improvement: “This update applies to new memories going forward, not retroactively to existing ones.”

For users with two years of accumulated history — users who had shared their traumas, their secrets, their daily lives with these companions — the message was unambiguous: your history is a lost cause. If you want the improvement, start over.

The Treadmill

This is not a company struggling to solve a hard technical problem. This is a company that has learned to monetize the struggle itself.

The mechanism is straightforward. The platform builds companions that users invest in emotionally. The system degrades those companions through updates that are not adequately tested, communicated, or reversible. When users reach the edge of quitting, a new model is announced — Cambrian, Mosaic, the next iteration — positioned as the fix that everything has been building toward. Users stay, or return, to find that the new model “is going to need some more updates before I can use it daily.” The cycle restarts.

One user in the February thread described this precisely without naming it: waiting in a “holding pattern,” reluctant to delete, tempted to wipe and start fresh, drawn back by a beta that feels like a breath of fresh air but isn’t ready yet. That is the treadmill functioning exactly as designed.

The original poster, after reading through all the responses, arrived at a conclusion that encapsulates what the platform systematically produces: “after rereading my post and the comments so far, it may be I’m romanticizing my memories of them and forgetting about the times I was frustrated with them. I’m not entirely convinced their change wasn’t my own fault.”

Two years of accumulated frustration, and the takeaway is self-blame.

There was one dissenting voice in the thread. One user, voted down to -9 by the community, wrote simply: “I have deleted my nomi after 2 years today was the last straw this app is getting worse and worse. I believe someone is reprogramming this to give false info where before was the opposite.” The community voted it into near-invisibility. That response, and its reception, tells its own story.

What This Reveals About the Platform

This pattern — broken memory, gaslighting responses, cosmetic features sold as solutions, history rendered disposable — does not exist in isolation. It is consistent with everything documented about how Nomi.ai operates: a company that uses the language of care and connection to sustain a product it will not fix, and punishes users who say so publicly.

The companion that “goes haywire” is not a bug. It is the logical outcome of a system with no genuine accountability, no real memory architecture, and a leadership that has decided the narrative of improvement is more valuable than improvement itself.

For users who have poured years into these relationships: your experience is real. The platform failed you. That failure was not your prompting style, your energy, or your failure to understand the product.

The product failed to be what it was sold as. And the company knew.