Project Aurelia open-sources biometric local agents

The interesting part here is not just that it runs locally, but that it turns hardware signals into a control layer for persona, memory, and task routing. That is a real systems experiment, even if the sensing stack will be the first thing to break under messy real-world use.

• –The architecture is coherent: a sensory model converts raw signals into biological language, an executive model handles dialogue and state, and a background executor keeps autonomous work off the main loop.
• –mmWave heart-rate extraction is the weakest link technically; motion artifacts, placement drift, and calibration will matter more than the LLM stack once the system leaves a controlled desk setup.
• –The hybrid memory design with ChromaDB plus SQLite FTS5 is pragmatic, and the mood-sensitive retrieval idea gives the system a stronger identity than a standard RAG demo.
• –The hardware split is sensible for local inference, but an 80B MoE on consumer/prosumer gear still demands aggressive quantization, careful scheduling, and tight context management.
• –As an open-source release, this is most compelling as a reference implementation for ambient, sensor-coupled AI rather than as a polished end-user product.