Artificial Local Intelligence: Architecture and Reference Implementation

Autonomous agents operating over extended periods face requirements that are qualitatively different from those of isolated task execution. A system that must continuously preserve operational integrity while generating and evaluating behaviour cannot be organised around a single reasoning process without sacrificing transparency and controllability. We present Artificial Local Intelligence (ALI), a reference architecture that decomposes autonomous operation into five explicitly separated responsibilities: a Causal Core that evaluates operational viability before planning begins, an Ego that generates behavioural proposals without evaluating them, a Super-Ego that evaluates proposals without generating them, a Memory that preserves every decision as a complete and immutable operational episode, and a Runtime that coordinates the cycle without participating in reasoning. We provide a complete architectural specification, a compliance criteria set of 44 verifiable SHALL requirements, and an open-source reference implementation in Python that requires no third-party dependencies. The implementation is verified by 72 automated tests and includes a plugin system that allows any architectural component to be replaced through configuration. A comparison against BDI architectures, SOAR, LangGraph, and the Microsoft Agent Framework shows that no existing production framework provides an independent architectural component for operational viability monitoring or enforces the separation of behavioural generation from normative evaluation at the architectural level.