Hermes Agent Self-Evolution Tested: After One Week, It Automatically Learned Your Work Habits — The More You Use It, the Better It Knows You
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
Abstract: Testing Hermes Agent's core black technologies: automatic skill evolution generates Markdown Skill files after task completion; multi-layer persistent memory uses SQLite + FTS5 full-text search + LLM auto-summarization for cross-session preference retention; deep user modeling via Honcho builds a personalized model.
I've been using Hermes Agent for a week now.
The version I had on Day 1 is completely different from the one on Day 7.
Day 1: It Asked About Everything
"What tone would you like me to use?" "Markdown or plain text?" "Where should I save this file?"
Every single task required confirmation. Annoying.
Day 3: It Started Making Its Own Judgments
"Based on your previous preferences, I wrote this in concise Chinese style, Markdown format, saved to your workspace directory. Let me know if you'd like adjustments."
It stopped asking. It made the call itself, then asked "want me to change it?"
Where did this change come from? It "remembered."
Automatic Skill Evolution: It Writes Skills After Completing Tasks
This is Hermes's most surprising capability.
I asked it to write a product description. After finishing, I casually mentioned: "Don't start with an exclamation mark next time. I don't like that."
The next day, I asked it to write another description. No exclamation mark at the beginning.
I checked its Skills directory — there was a new Markdown file: writing-style-preference.md. The content was self-generated: "User prefers concise Chinese style, no exclamation marks at the beginning, blank line between paragraphs."
It didn't just "remember" what I said. It abstracted my preference into a reusable skill file.
Here's the mechanism: every 15 tool calls, Hermes triggers a self-iteration. It analyzes those 15 interactions, extracts patterns and preferences, and generates or updates a Skill file. Next time it encounters a similar task, it loads that Skill automatically.
You don't need to manually configure anything. The more you use it, the better it understands you.
Multi-Layer Persistent Memory: No Forgetting Across Sessions
Most AI assistants forget everything when you close them. What you told them yesterday is gone today.
Hermes's memory has three layers:
Short-term memory — context within the current conversation. Every AI has this.
Medium-term memory — SQLite database + FTS5 full-text search. Everything you've said, done, and expressed preference for is stored. Next time you start a new session, it can search historical records. Not a fuzzy "maybe remembers" — precise full-text retrieval.
Long-term memory — LLM auto-summarization. Hermes periodically compresses medium-term memory into summaries, forming your "user profile." For example: "This user is in content operations, prefers concise style, commonly uses WeChat and Feishu, dislikes AI-sounding writing."
Three-layer memory in action: short-term handles immediate problems, medium-term handles cross-session continuity, long-term handles deep personalization.
Honcho User Modeling: Your Personal AI
Hermes's third layer of innovation is Honcho — a deep user modeling system.
Simply put, Honcho builds a dedicated small model for each user. This model doesn't replace the large model — instead, it applies a layer of personalization before the large model reasons.
For example, when you ask Hermes to "write an email": - Without Honcho: Hermes writes a standard email using a generic template - With Honcho: Hermes first queries your user model, discovers you prefer direct business style, habitually use informal greetings, and like adding an English quote to your signature — then writes according to your style
This is the underlying mechanism of "the more you use it, the better it knows you." Not magic — data.
One Week of Changes
| Day 1 | Day 3 | Day 7 |
|---|---|---|
| Asks about format every time | Judges format itself, just confirms | No format questions, executes based on preferences |
| Generic writing style | Mimics my previous style | Writes like I wrote it myself |
| Doesn't know my preferred tools | Remembers my 3 most-used tools | Proactively suggests better tool combinations |
| Waits for instructions after tasks | Proactively suggests next steps | Anticipates my needs, prepares in advance |
By Day 7, I asked it to "organize today's topics." It didn't just organize them — it sorted them by my previous priority logic, and flagged similar topics I'd already written about to avoid duplication.
Once you've experienced this, going back to a regular AI assistant feels like something's missing.
Runs Locally, All Data Stays With You
All memory data — Skill files, SQLite database, Honcho user model — is stored locally.
AI Box (also known as Agent Computer or AI Box) is a dedicated local hardware device that runs AI Agents, pre-installed with an AI agent management system, plug-and-play, running 24/7. Hermes runs on Kaihe AIBOX — your preferences, habits, and work patterns all stay on the local device, never uploaded to any cloud server. You control it remotely via WeChat or Feishu, and the Agent's "understanding you" comes from local memory systems, not cloud data mining.
Want to Go Deeper?
Official Website (agentaibox.com) — local Agent that gets smarter the more you use it "Hermes Agent v0.12.0 Architecture Revolution: Kanban Multi-Agent Collaboration" — multi-Agent architecture "This AI Takes Notes — After a Month, I Found It Gets Smarter Every Day" — persistent memory experience
-#KaiheAIBOX #Hermes #24x7Online #AIAgent #LocalAI
Kaihe AIBOX | The Agent Computer That Works 7×24 for You · Hermes Zone