Can a 4B local model actually feel like an AI assistant?

I've been building Arcon around Qwen3-4B + LoRA. Instead of just making it a chatbot, I'm experimenting with persistent memory, personality/mood, internal state, tools, and eventually having it process things before replying.

I'm curious what people who've built local agents think - how far can you realistically push a small model with good architecture around it?

I put the whole thing on GitHub if anyone wants to poke around, roast the architecture, or tell me what I'm doing wrong, stars are always appreciated!