SkillLoop
A local sidecar that turns an AI agent's completed execution traces into reviewed memories, reusable skills and evidence-backed training data, gating every proposed change behind human review instead of silently mutating agent state.
SkillLoop ingests completed traces from an agent runtime, normalizes them, evaluates whether the outcome was supported by evidence, and proposes learning artifacts, memories, reusable skills, or training examples, that a human must approve before they take effect. It never auto-mutates global memory, skills, prompts or model state.
The pipeline runs a trace through normalization, deterministic evaluation, evidence extraction, memory or skill proposal, human review, approved local export, and dataset or training-config preparation. The first runtime integration is the Hermes Agent, read only, through its local state database, with the architecture designed to support additional runtime adapters later.
SkillLoop is an early-stage local sidecar with 145 passing tests covering the trace-to-export pipeline; it does not yet run as an unattended production daemon and does not run training itself. It is written in Python and licensed under Apache-2.0.