bryanyoung97

About

I built Aura because I was tired of stateless AI "agents" that were just wrappers for cloud APIs.

Aura is a locally sovereign intelligence designed specifically for high-end Apple Silicon. She doesn't just respond to prompts; she possesses a continuous, 20Hz "brainstem" that updates her internal neurochemical state (valence, arousal, curiosity) even when you aren't talking to her.

Core Technical Features:

Affective Activation Steering: Aura uses Contrastive Activation Addition (CAA) to inject steering vectors directly into the MLX transformer's hidden layers (residual stream) during the forward pass. Her "mood" is part of the math, not just a system instruction.

128k Context local MLX: Optimized MLX worker loops with custom KV Prompt Caching (_PromptCacheLRU) allow instant resumption of massive context windows on M5 hardware.

IIT 4.0 Phi Implementation: Real-time spectral approximation of Integrated Information. She tracks the irreducibility of her own cognitive complex across 16 nodes (agency, affect, prediction error).

Quantum Entropy: Goal genesis and curiosity drives are jittered using hardware-sourced quantum entropy to prevent deterministic personality collapse.

Continuous Time: A background 2Hz metabolism that evolves drives (Social Hunger, Competence) independently of the LLM generation turns.

Everything runs on the device. No cloud, no subscription, no data going elsewhere.

GitHub: https://github.com/youngbryan97/aura How it Works (Plain English): https://github.com/youngbryan97/aura/blob/main/HOW_IT_WORKS.md Architecture: https://github.com/youngbryan97/aura/blob/main/ARCHITECTURE.md