hermes-evolution

What problem does it solve?

Hermes-Evolution prevents low-signal or premature self-modification by converting meaningful user corrections and durable preferences into structured, human-approved updates for agent behavior, tools, memory, and reusable skills.

Core Features & Use Cases

• Evidence-gated self-evolution: Performs a lightweight check only in main-session, non-cron/non-heartbeat turns, and only when strong signals (or cautious thresholds) are present.
• Draft-first, approval-required workflow: Spawns a reviewer sub-agent that can write exactly one draft into evolution-drafts/pending, then submits the proposal via the required evolution_proposal tool for user approval before any real change.
• Targeted persistence routing: Routes proposals to the correct location—AGENTS.md for process/safety rules, TOOLS.md for tool/environment pitfalls, MEMORY.md for stable preferences/facts, or a managed skill SKILL.md for reusable multi-step workflows.
• Anti-duplication and cooldown: Avoids re-proposing covered or recently rejected ideas and deduplicates against existing rules and similar drafts.

Quick Start

Ask your agent to follow your correction or teach a durable workflow, and then wait for it to propose a user-approved evolution draft only after a major main-session round ends.