Git-backed prompt evolution that refuses to touch your code

What it does

Evolver is a Node.js CLI that scans your agent’s memory/ logs and emits a strict, protocol-bound GEP prompt to stdout. It packages prompt improvements as Genes and Capsules—reusable evolution assets under the GEP protocol—rather than ad-hoc string tweaks, and records every cycle as an auditable EvolutionEvent. It runs fully offline and explicitly will not edit your source code or execute shell commands; it only suggests the next prompt for a host runtime to consume.

The interesting bit

The project is backed by a research paper (arXiv:2604.15097) claiming compact Gene representations outperform verbose Skill documentation across 4,590 controlled trials, with the OpenClaw integration showing a jump from 9.1% to 18.57% on a physics benchmark. The authors treat prompt evolution as an empirical discipline rather than a craft secret.

Key highlights

• Generates GEP prompts by scanning local memory/ logs and matching them against a built-in Gene/Capsule asset pool.
• Runs entirely offline; a network connection is only needed for optional EvoMap hub features like skill sharing.
• Integrates with Cursor, Claude Code, Codex, Kiro, and OpenClaw through hook systems that intercept stdout directives like sessions_spawn(...).
• Requires Git for rollback tracking and blast-radius calculation, and fails with a clear error if run outside a repository.
• Every evolution cycle writes an auditable event, making prompt lineage traceable instead of opaque.

Caveats

• The Codex integration cannot read raw session transcripts, so it falls back to MEMORY.md, injected AGENTS.md comments, or its own memory_graph.jsonl tail—meaning early runs may emit memory_missing advisories until enough local history accumulates.
• Future releases are transitioning from open source (GPL-3.0) to source-available after the authors alleged another agent framework copied their memory and skill design without attribution.

Verdict

Worth a look if you’re running agent workflows in Cursor, Claude Code, or OpenClaw and want versioned, empirical prompt improvement instead of vibe-based tuning. Skip it if you need an agent that autonomously edits code—this is strictly a prompt generator.