NPCs that remember you and talk back
October 2026
Talking to a game character in your own words already works. I wanted to know what’s underneath, because I want to build it in Unreal.
It’s already in real games. inZOI runs a tiny model on the player’s own GPU for its “Smart Zoi” characters. In Whispers from the Star you keep a stranded astronaut alive just by talking to her.
Local or online?
Most guides end up suggesting a mix: cloud for the big moments, local for the everyday conversations.
How does it remember?
The model never learns anything. The game just puts the right memories in front of it every turn. They’re saved in three layers: the last few messages word for word, older stuff summarized in a vector database, and a facts table for things it must never forget.
day 2 “My sword broke in the caves.”
day 5 “Can you fix that?”
npc “Is that the sword from the caves?”
The game found the closest saved memory, added it to the prompt, and that’s all it took.
In Unreal
- NVIDIA ACE: runs locally and shares the GPU with the graphics.
- Convai: easiest to start, Blueprints only, but online.
- llama.cpp plugins like Volley Studios’ Local Multimodal LLM: local, free, still in beta.
The one rule every guide repeats: the AI can suggest an action, but the game decides if it happens.
What still breaks
Delay, memory that quietly stops working, characters slipping out of role, and players tricking NPCs into spilling their secrets.
Sources
Sometimes I zone out and just think about everything that’s happened to me.
How can you do everything right and still feel like you’re not doing enough? And the results of what I do always seem to go against me.
I don’t really know. It’s hope that keeps me up at this point. Hope that someday everything will be fine and I’ll figure it out.
And then I ask myself what the point of trying is anyway.