I spent the last few months building RizzMaster, an AI dating simulator where you text AI characters in realistic conversations. The hard part was never the chat. It was making characters feel like people instead of a generic assistant. Here is what actually moved the needle.
Memory is the whole game
A chatbot that forgets you is a toy. The moment a character brings up something you said 20 messages ago, it stops feeling like a bot.
What worked for me:
- Summarize the conversation into compact running state instead of stuffing the full history into context every turn. Cheaper and it scales.
- Keep short-term memory (this session) separate from long-term memory (facts that persist across chats).
- Store extracted facts as structured data, not raw transcript. "likes horror movies" is more useful than re-reading the whole thread.
Personality has to be consistent, not random
Early on every character drifted into the same polite voice. Two things fixed it:
- A fixed persona spec per character (traits, speech style, boundaries) injected every turn.
- Let characters say no. One that always flirts back is boring. Ones that go cold, get bored, or block you when you are rude feel real.
Behavior beats word choice
The thing testers reacted to most was not clever lines. It was reactions. Characters that text first. Characters that ghost. Trust that moves up and down based on how you actually talk. Behavior is what sells "this is a person".
Takeaways
- Persistent memory is the single biggest lever for making AI feel human.
- Give characters the ability to react negatively. Constraints create personality.
- Structured long-term memory scales far better than dumping history into context.
It is live if you want to see the ideas in action: RizzMaster. Happy to go deeper on the memory setup in the comments.
Top comments (1)
If you want to try the actual app, it is on the App Store here: apps.apple.com/app/id6757169407. Happy to dig into how the memory and persona system works if anyone is curious.