I Is Not Singular builds 4 LoRA agents

This is a strong proof-of-concept for weight-level personalization on consumer hardware, not just prompt cosplay.

• –Per-agent LoRA solved the “majority persona wins” problem that showed up with a single shared adapter
• –The memory math is the real enabler: Q4_K_M base plus four small adapters stays inside 8GB, and training still fits
• –The split between persona LLM and shared inner modules is the most interesting part; it gives a cleaner model of identity than prompt stacks do
• –The sleep-cycle retraining loop makes the agents’ evolution observable and grounded in their own experience instead of external steering
• –The practical stack notes matter too: CUDA-built llama.cpp, hot-swappable adapters, and multilingual cleanup are the difference between a demo and a working system