The learning loop

This is what makes Hermes fundamentally different from a chatbot: a closed loop where the agent gets smarter over time.

The loop

        ┌──────────────┐
        │  You give    │
        │  a task      │
        └──────┬───────┘
               ▼
        ┌──────────────┐
        │  Hermes acts │ ← tools, files, shell
        └──────┬───────┘
               ▼
        ┌──────────────┐
        │  Outcome     │
        └──────┬───────┘
               ▼
   ┌───────────────────────┐
   │  Memory updated       │ ← facts, preferences persist
   │  Skill created/refined│ ← procedures packaged for reuse
   │  User model deepens   │ ← learns how you work
   └───────────────────────┘
               │
               ▼
        Next task starts smarter

Four components

1. Agent-curated memory: Hermes decides what's worth remembering and stores it in ~/.hermes/memory/.
2. Autonomous skill creation: after repeated tasks, Hermes packages the procedure as a reusable skill.
3. Skill self-improvement: skills get refined during use, not just at creation time.
4. Cross-session recall: FTS5 search + LLM summarization pulls relevant past context into new sessions.

What this means in practice

Week 1: You explain your project structure every session.

Week 3: Hermes remembers your stack, your conventions, and how you like files organized. You skip the briefing.

Week 6: Hermes has a skill for "update the VitePress sidebar when I add a new page." You say "add a page about X" and it handles config + content + sidebar.

How to accelerate the loop

• Be consistent. Work from the same project directories so context files and memory align.
• Correct mistakes. When Hermes gets something wrong, tell it explicitly. Good corrections become memory.
• Repeat tasks deliberately. The second and third time you do something, Hermes is more likely to skill-ify it.
• Use SOUL.md and context files. Static context frees memory for dynamic learning.

Offline advantage

With local models, the learning loop is entirely on your hardware. Memory and skills in ~/.hermes/ never touch a cloud API. Your agent's "brain growth" stays private.

Next: Memory system and Skills system.