Xingye Onboarding: Great Roleplay, Strict Boundaries

I usually spend my time running APIs and agents (automated task executors), so I rarely use role-companion apps seriously. This week I tried Xingye. The homepage features recommended characters, agent creation, chat entry points, plus voice and memory photos. I first chatted briefly with a pre-made character, then created my own lab assistant, filling out the persona, selecting a voice, and setting the tone. The whole process didn't require writing prompts, which is beginner-friendly.

The pros are clear. It lowers the barrier significantly: avatars, voices, personas, and plot entry points are all made into forms—you click a few times and get a chat-ready agent. Feedback is also complete; during chats, you receive memory photos and Star Thoughts.

Star Thoughts are AI-generated character photo cards that turn conversation results into visual rewards, making you feel like the other party remembers you. It also supports User Generated Content (UGC) distribution; characters can be shared, and Star Thoughts can be listed for sale. In my testing, quickly trying out a character shows their packaging capabilities are indeed mature.

But the pitfalls are direct, and the most glaring is content boundaries. I originally thought the difficulty in role companionship was acting convincingly, but in reality, I got stuck on where the conversation would be blocked. I tried a few ordinary plot segments; as soon as the relationship got slightly closer, the dialogue started dodging, as if the system had drawn a warning line in advance. App store reviews also mention complaints, saying even seeing abs is sensitive. This restriction isn't necessarily wrong, but it interrupts the narrative. Just as the atmosphere builds up, the character suddenly turns back into customer service, and immersion drops sharply.

Voice has issues too. Xingye sounds okay, but in my tests, text and voice emotions often mismatch. A sentence that should be slow and soft might suddenly speed up; a response that should be firm might have a wavering tail. This looks like the model understood the character on the text side but didn't pass the full emotional state to the voice side. Research on character consistency often discusses persona drift; Xingye's voice issues expose this more clearly.

Long conversations are unstable. In short chats, the character feels human, but after ten-plus rounds, the settings start leaking. I asked it to remember a detail, and later it forgot; I asked it to maintain a certain tone, and after a while it turned into a generic assistant. It has memory and photos, but I can't tell what actually entered the context versus what's just for display. For regular users, this is just a minor forgetfulness. For those seriously doing roleplay or researching models, this is unreliability.

Monetization affects the experience too. Star Thoughts have free gacha pulls, after which you need Star Diamonds. This design is great for retention but turns companionship into a task. You're chatting, and suddenly realize there are gachas, backgrounds, selling, and revenue sharing. It's lively, but the boundary is slippery—are they providing emotional connection or driving consumption?

If you want to experience AI character chat, don't want to fuss with prompts, and can accept moderation and gacha mechanics, give it a try. It's also good for observing UGC sharing and transaction chains in consumer products. If you want stable long plots, highly consistent characters, auditable memory, or are sensitive to content boundaries, don't expect too much. Xingye can chat. The wall comes fast.