Chapter 15: Hermes Agent — The Self-Evolving Agent
☤ "The longer you use it, the more it understands you — not marketing, a closed loop."
Chapter Introduction
The previous chapter showed OpenClaw — a "consumer-grade" personal assistant living in chat apps, where all capabilities are maintained by humans. Hermes Agent takes a different road: the Agent writes its own skills.
Hermes Agent (nicknamed "爱马仕" by the Chinese community) was released by Nous Research in mid-2025. It shares deep lineage with OpenClaw (Node + TypeScript, personal local execution, multi-platform), but its core differentiator is one word: learn.
- OpenClaw is "configure-and-use" — skills are pre-written by users/community;
- Hermes is "learn-in-use" — after each complex task, the Agent itself distills the execution trajectory into a Markdown skill file (
SKILL.md) and iterates those skills offline (DSPy / GEPA-inspired).
This "autonomous skill creation + continuous self-iteration" is called Self-Evolving Skills — the sharpest feature distinguishing Hermes from every other Agent framework. Combined with a three-layer memory system and a proactive Nudge Engine, Hermes is "the agent that grows with you."
This chapter:
- Dissects the closed learning loop — how "write skills automatically + improve skills automatically + reflect proactively" forms an engineered self-evolution loop;
- Demonstrates install / migration —
hermes claw migrateone-click migration from OpenClaw; - Analyzes three-layer memory + FTS5 + LLM summarization — how SQLite full-text search + LLM summarization resolve "limited context vs. long-term memory."
Chapter Content Overview
| Section | Content | What You'll Learn |
|---|---|---|
| 15.1 Birth & the "Growing Agent" Philosophy | Nous Research, the naming, the self-evolution design | Hermes' place in the new-gen Agent spectrum |
| 15.2 Installation & Migration (from OpenClaw) | one-liner, hermes claw migrate | Run it on macOS/Linux/VPS/Termux |
| 15.3 Three-Layer Architecture | Gateway / Engine / Plugin subsystems | Read Hermes' engineering skeleton |
| 15.4 Core: Self-Evolving Skills Loop | skill distillation, DSPy/GEPA offline iteration | How "Agent writes its own skills" works |
| 15.5 Three-Layer Memory | long-term / working / episodic, FTS5 + LLM summary | Cross-session memory design |
| 15.6 Nudge Engine & Cross-Session Learning | active reflection, Honcho user modeling, cron | Make the Agent proactively ask "what to remember" |
| 15.7 Borrowing the Philosophy | closed loop, model-agnostic, runs on $5 VPS | Bring Hermes' philosophy to your system |
Reading Recommendations
- ✅ Engineers wanting to understand Agent self-evolution — focus on 15.4, 15.6
- ✅ OpenClaw users wanting to upgrade — focus on 15.2
- ✅ Long-term memory / cross-session developers — focus on 15.5
Prerequisites: Chapters 8 (Harness), 9 (Skill System), 14 (OpenClaw). This chapter and Ch.14 form the "action-Agent twin chapters" — one multi-channel, one self-evolving.
Citation Convention
Verifiable facts are anchored to the Hermes Agent main branch and hermes-agent.nousresearch.com/docs/. Time-varying numbers are not cited.
💡 Core insight: Hermes proves an Agent can distill experience into skills like a person — provided it's designed as a closed system loop, not maintained by hand.
Next section: 15.1 Birth & the "Growing Agent" Philosophy