GEAR: Git-backed Error Analysis & Reflection

GEAR (Git-backed Error Analysis & Reflection) is a protocol specification for AI agent error learning. It addresses a fundamental problem in agent systems: errors corrected in one session reappear in the next, because most memory architectures are session-scoped and lack quality control. GEAR solves this through three mechanisms: a Production-Audit-Consumption (PAC) model that separates rule creation, validation, and application into distinct roles; a Δ decision factor that routes each rule to automatic, semi-automatic, or mandatory human review based on confidence, risk level, and empirical evidence; and git-backed storage that makes reflection rules persistent, versioned, and atomically readable across sessions. The protocol defines five roles (Orchestrator, Resource Creator, Checker, Learner, Searcher), a five-state rule lifecycle, a structured YAML metadata schema with intent-driven retrieval, and nine conformance requirements. It is implementation-agnostic and requires no model fine-tuning. A reference implementation, Aristotle, demonstrates the protocol in a Claude Code / OpenCode environment. Source: https://github.com/alexwwang/aristotle