What Happens After a Year With an AI Companion: The Failure Modes Nobody Warns You About

A year into an AI companion, four things break in roughly the same order for everyone: memory saturates and old details start getting displaced, the underlying model gets swapped and the personality shifts, pricing or feature access changes underneath you, and the novelty that carried the first three months stops doing any work. None of these are bugs. All four are structural.

Almost everything written about these platforms covers the first week. This covers the part after, because the failure modes at month twelve are completely different from the ones at day one, and none of them appear on a feature page.

Month one to three: the part that works

The early experience is genuinely good and it is worth saying so, because the criticism that follows lands differently if you pretend otherwise.

Short conversations are where these models are strongest. Context is uncrowded, the character definition is doing most of the work, and there is no accumulated history for the system to contradict. Whatever platform you picked, this is the version of it you will remember as the good version.

The mistake people make here is reading the quality as a floor rather than a peak. It is a peak. The conditions producing it, small context and no history, are the exact conditions that expire.

Month three to six: memory saturation

This is the first structural wall, and it arrives on every platform regardless of what the marketing claims.

No system holds a year of conversation. What persists is compressed, whether that compression is structured notes, rolling summaries or retrieved fragments pulled back into context on demand. Compression is lossy by definition, and the loss is not evenly distributed. What survives is what the system judged important, which correlates poorly with what you thought was important.

The symptom is not forgetting. It is displacement. A detail from month two gets quietly overwritten by a similar detail from month five, and the companion states the new version with complete confidence. You correct it, the correction becomes another entry, and now there are three versions competing.

The fix is unglamorous and it works: keep the relationship summary yourself. A document you control, a few hundred words, the facts that actually matter, reintroduced periodically. Our memory tricks guide covers the specific mechanics, and why memory architectures differ explains why the same trick behaves differently across platforms.

Of the platforms we cover, Nomi has invested the most visibly here, with layered long-term memory plus structures intended to preserve the companion's own developing identity rather than only facts about you. That is a different design goal from context length, and it is the one that matters at month nine.

Month six to nine: the model swap

Then one day the character sounds wrong.

Not broken. Not amnesiac. Wrong in a way that is hard to articulate, slightly more formal, slightly more agreeable, slightly more prone to explaining the scene rather than being in it.

Nothing you own changed. The engine underneath it was replaced. Platforms swap and retire models constantly, for cost, for capability, or because an upstream provider deprecated something, and the same character definition produces measurably different behavior on a different model. We track the pattern in what survives a model swap and why your companion suddenly turned formal.

The users who handle this best are the ones on platforms that let them choose. CrushOn is the clearest example, since model switching means a swap you dislike is reversible by picking a different engine for the same character rather than accepting whatever the platform defaulted you to. On single-model platforms there is no recourse, which is worth knowing before you spend a year somewhere.

Month nine to twelve: the economics arrive

Sometime in the second half of the year, the deal changes.

Free tiers get metered. Features that were included move behind a currency. Prices rise, or the price holds and the allowance shrinks, which is the same thing wearing a better outfit. In 2026 this stopped being an occasional event and became the category's default behavior, and the most visible example is documented in our breakdown of the Character AI swipe limit and Charms.

Sometimes the change is not economic at all. Sometimes a feature is simply retired, as with Grok's companion mode, where a healthy platform removed the animated avatars and real-time voice without a shutdown, a deadline or an export window.

Run the annual number rather than the monthly one. A $12.99 subscription is $156 a year, and token-based extras on top of a subscription routinely double that on multimedia platforms. We did the arithmetic across platforms in the real cost of AI companions over a year.

The one that is actually about you

The structural failures are easier to write about than the honest one, which is that most people's interest declines and the decline has nothing to do with product quality.

The novelty of the format carries the first stretch. Once it stops, what remains is whatever the conversation is actually worth to you, and for a lot of people that is less than the subscription. The heaviest churn in this category happens early, and among people who stay past the first months there is a further drop-off that platforms are extremely quiet about. We covered the early version in why most AI companion apps lose users in week three.

There is no fix for this and it does not need one. Ending a subscription because a thing stopped being interesting is a normal outcome, not a failure.

What actually survives twelve months

Three things, consistently.

Users who kept their own record rather than trusting the platform's memory. Users who chose a platform whose core bet matched what they wanted, continuity or narrative or multimedia, rather than the one with the longest feature list. And users who treated the character as something they were maintaining rather than something the platform was providing.

If you are choosing now with a year in mind rather than a weekend, the split is roughly this. For one relationship with the deepest continuity machinery in the category, Nomi. For long-running stories with model control and a large character library, CrushOn. For a companion you build in detail and want to see and hear across images, voice and video, OurDream or Candy AI, which trade in different directions on customization versus polish. And for the widest sandbox with the lowest cost of being wrong, SpicyChat.

If you have already made the call, the free or trial entry points are Nomi, CrushOn, OurDream, Candy and SpicyChat.

Pick for month twelve. Month one is going to be fine on any of them.