Aegis Falls is my personal AI architecture. The name captures the flow: frontier intelligence falling from the top of the stack down into local execution. Claude runs on OpenClaw at the apex, handling reasoning and planning. Tasks cascade from there into smaller local agents running on-device via Ollama. Lily is the AI persona — the persistent identity that threads the whole system together.
Claude handles the hard thinking. Local agents handle the doing. Lily is who you're actually talking to. The cascade is the architecture; the fall is the flow of intelligence from frontier model to local execution. Everything runs on hardware I own — a dedicated GPU node (aegis-node, AMD Radeon AI PRO R9700, 32GB VRAM, ROCm) with a secondary embedder/inference node for background tasks. The frontier is there when you need it. Local is always on.
What makes it interesting is the ACT-R cognitive memory layer — a persistent memory system inspired by human cognition. Information isn't just stored; it has activation levels based on recency and frequency of use. Retrieval works through spreading activation, meaning related memories strengthen each other. It's not a simple database lookup — it's a cognitive architecture that learns which information matters.
The whole system runs 24/7. Lily handles tasks autonomously through Discord, manages the homelab, assists with development, and maintains context across sessions. It's not a product — it's a personal tool, built iteratively over six versions, running on hardware I own. Read the full architecture deep-dive.