Adaptive Runtime – AI agent layer for production resilience without GPU, offering crash recovery and state persistence via SQLite

A runtime intelligence layer (not an LLM framework or workflow builder) that adds crash recovery, state persistence, confidence scoring, and contextual adaptation to AI agents, designed to run on minimal hardware including $5 VPS instances.

Five core engines work together: State Engine (persistent memory via SQLite), Context Engine (transforms raw signals into risk/stability/pressure scores), Confidence Engine (adaptive probabilistic scoring with historical decay), Decision Engine (rule-based action selection), and Recovery Engine (checkpoint snapshots and exponential back-off retry).

Requires only Python, asyncio, pydantic, and aiosqlite — no GPU, no cloud lock-in, and designed for constrained environments including Raspberry Pi, edge devices, and air-gapped systems.