What AI Girlfriend Apps Still Get Wrong, Two Years In

Giving credit where it's actually due first

Before I get critical, I want to be fair about what's genuinely improved. Chat quality across the 129 platforms I test now averages 3.26 out of 5, a solid number for a technology that's still fundamentally young. Video generation went from nearly nonexistent to 22% adoption in a short window. Pricing has gotten more accessible, with 48% of platforms now offering a genuine free tier. This is a category that has made real, measurable progress, and I don't think it's fair or accurate to frame it as stuck.

But "has made real progress" and "has solved its hardest problems" are different claims, and I think this category has quietly let the second one slide while making genuine headway on the first. That gap is what this article is actually about.

What still gets memory wrong

Memory is the clearest example. Only 21% of platforms document a real cross-session memory system, meaning most AI girlfriend apps still effectively reset or lose track of what you've told them between sessions. I've covered the data behind this specific number in more depth in a dedicated piece on why cross-session memory is rarer than you'd think, and my honest opinion, two years into watching this category mature, is that this number should be a lot higher by now given how consistently users cite memory as the thing they want most from a companion app.

What frustrates me about this specific gap is that it's not a new problem the industry just discovered. It's been the most requested, most obviously unsolved feature for as long as I've been testing this category, and the pace of improvement here has been genuinely slow compared to how quickly video generation and pricing have shifted in the same window.

What still gets voice wrong

Voice interaction is the second problem, and it's arguably the most visible one. It averages just 1.81 out of 5 across the 129 platforms I test, meaningfully lower than every other category I score, and 77% of platforms lack functional voice interaction entirely. In my opinion, this reflects an industry that keeps treating voice as an afterthought feature to bolt on, rather than core infrastructure worth investing in properly. A platform that leads its marketing with voice and then delivers a mediocre, laggy, or robotic-sounding version of it is, honestly, worse off than one that just stuck to text.

I don't think this is purely a technology limitation at this point. The underlying speech technology has improved a lot industry-wide over the past couple of years, in general-purpose AI assistants far beyond this specific category. The fact that AI girlfriend apps specifically haven't kept pace with that broader improvement suggests to me it's more of an investment and prioritization gap than a pure technical ceiling.

What still gets support wrong

The third problem is the one I think gets the least attention relative to how basic it is: customer support. 78% of the platforms I track have no documented support channel at all. That's not a hard AI research problem the way memory and voice are. It's a basic operational gap, and in my opinion it's the most inexcusable of the three precisely because it doesn't require any technological breakthrough to fix, just genuine investment in a part of the product that doesn't show up in a demo or a screenshot.

I've written more about the specific pattern of what users complain about across the reviews I've collected in a dedicated pattern analysis of AI girlfriend app complaints, and support gaps show up there constantly. Two years into this category's real growth, I'd have expected basic support infrastructure to have matured faster than it has, especially given how personal and potentially sensitive the product itself is.

Why the average score still lands at 2.5 out of 5

These three unsolved problems are a big part of why the overall average across 129 platforms still sits at 2.5 out of 5, a middling score for a category that's otherwise clearly grown up in other respects. I've written specifically about why we score the industry this way, and why we don't grade on a curve, in a dedicated piece explaining that score and our methodology behind it. My opinion is that the 2.5 average isn't a sign this category hasn't grown. It's a sign that growth has been lopsided, real progress on the easy, visible features and real stagnation on the hard, foundational ones.

My honest opinion on why this pattern persists

I think the honest explanation is incentives, not inability. Chat quality, images, and video are all things a company can demo convincingly in a screenshot or a short clip, which makes them easier to market and easier to justify investing in from a growth perspective. Memory, voice quality, and support don't demo well. They only become visible over weeks of actual use, or in the moment something goes wrong, which makes them easy for a growing company to underinvest in relative to how much users actually care about them once they're using the product for real.

That's not a moral failing on any single company's part. It's a predictable pattern in a fast-growing, still-consolidating market with 129 competitors, where the companies rewarded fastest by growth metrics are often the ones optimizing for what's easiest to show off, not necessarily what matters most for long-term user trust.

A smaller, recurring complaint that's really the same problem wearing a different hat

Beyond the three big gaps, I keep running into a smaller, related complaint across the reviews I collect: characters that feel inconsistent, saying or reacting in ways that contradict what was established earlier in the same conversation, let alone across sessions. I think this is really the memory problem wearing a slightly different hat. A model without a reliable way to track what's already been established about a character or a relationship is going to occasionally contradict itself, and no amount of chat-quality polish fully papers over that underlying gap.

I flag this specifically because it's easy to file "inconsistent personality" and "forgets things between sessions" as two separate complaints when they're really pointing at the same root cause. Fixing memory properly would likely resolve a meaningful share of this smaller complaint too, which is one more reason I keep coming back to memory as the highest-leverage fix available to this industry right now.

What I think would actually fix this, two years from now

• Memory needs to be treated as core infrastructure, not a feature flag, given how consistently it's the top complaint I encounter across reviews.
• Voice needs real investment, not a checkbox feature, since a mediocre voice implementation actively hurts a platform's perceived quality more than having no voice feature at all.
• Support needs to exist at all, full stop, before it needs to be excellent. Right now the bar is simply "does a documented channel exist," and 78% of platforms don't clear even that.
• Reviews and rankings need to keep weighting these categories heavily, so platforms that skip them don't get rewarded purely for having flashier visuals. Our best AI girlfriend ranking is built specifically to weight memory, voice, and support this heavily.

My bottom line, in my own opinion

Two years into watching this category mature, I think the honest verdict is: real progress, real remaining gaps, and the gaps are concentrated in exactly the features that would make this feel like an actual ongoing relationship rather than an impressive demo. I don't think that's a reason to write off the category. I think it's a reason to keep pushing platforms to invest in the unglamorous fundamentals as hard as they invest in the next flashy feature. I go into where I think this specific tension is headed next in my broader look at where the AI girlfriend industry is actually headed, and you can read more about how I test and score every platform or my background as a researcher in this space.

Further reading

• What Users Complain About Most in AI Girlfriend Apps: A Pattern Analysis From Our Reviews
• Why the Average AI Girlfriend App Scores 2.5/5 (and Why We Don't Grade on a Curve)
• The best AI girlfriend apps, ranked