Project: NousResearch/hermes-agent · Author: Nous Research · License: MIT
I forked the repository to read the Agent Loop, Skill Manager and Memory implementations. I did not contribute upstream.

What it is trying to solve

Hermes Agent is an open-source general-purpose agent released by Nous Research in February 2026. Unlike a single-turn chatbot, it focuses on long-running operation: remembering previous tasks, turning experience into reusable skills and being reachable from different messaging platforms.

It competes with OpenClaw; it is not a successor. OpenClaw broke out earlier, and Hermes arrived a few weeks later with a stronger emphasis on self-improvement and deep customisation.

Core mechanisms

Closed-loop learning

1
2
3
task → experience log → skill creation → skill refinement → persistent knowledge
  ↑                                                              ↓
  └──────────────── later tasks call the skill ←─────────────────┘

The flowchart is not the interesting part. These three questions are:

  1. Under what conditions does experience become a skill?
  2. How does a new skill avoid conflicting with existing ones?
  3. Who verifies the quality of a generated skill?

Without answers, “self-evolution” can easily become “adding bugs to itself”.

Memory

  • SQLite + FTS5 full-text search
  • Automatic memory summaries
  • Periodic “memory recall”
  • Cross-session continuity

The biggest risk in a memory layer is not forgetting — it is misremembering and remembering too much. The former contaminates later decisions; the latter raises both context cost and noise.

Platforms and automation

Telegram, Discord, Slack, WhatsApp and CLI are all supported. A built-in Cron scheduler can run daily reports, health checks and backups unattended.

My take

Hermes points in the right direction: an agent should be a system that accumulates, not a one-off Q&A. But it also moves the hard problem from “can the model do it” to “how do you keep it under control” — skill permissions, memory cleanup, cost control and rollback matter more than the demo.

View original project