AI Coworkers: Not Just Chatbots or Agents

Hermes Is Now the Most-Starred Agent on GitHub

Hermes Agent from Nous Research has become the most-starred agent framework on GitHub: 222.6k stars, ranked #20 among all repositories ever, with 42.7k forks and 2.5k contributors. Three things strike me about that:

  • What it says that a memory-first design overtook the gateway-first one
  • The migration path that got it there
  • The repo sitting one rank above it

The learning loop won

I covered Hermes in the six-framework survey: it is built around memory, storing sessions in SQLite, searching them with FTS5, distilling skills from experience, and building a persistent model of its user. Its tagline is “the agent that grows with you”.

OpenClaw got there first with the opposite bet: breadth of channels and the biggest skill registry. The comparison coverage frames the split as agent-first versus gateway-first, and by star count the field has picked the memory side. That matches my read of where the pain sits. Connecting an assistant to Telegram is a solved problem, while an assistant that greets you as the same stranger every morning is the unsolved one. The unsolved problem is accumulating useful state, and Hermes made that the core loop rather than a feature.

For ai-coworkers I took Hermes’s storage answer almost verbatim, FTS5 over SQLite plus distilled markdown, and pointed it at a different question: not “what did my user teach me” but “what does my role require”.

The migration path did real work

Two adoption mechanics stand out in how Hermes grew. hermes claw migrate moves an OpenClaw installation across automatically, skills, memories, and settings, which let OpenClaw users defect in one command. And Hermes adopted the agentskills.io portable skills standard plus support for installing ClawHub skills directly, so switching cost approached zero: your accumulated customisation came with you.

That is the same lesson mattpocock/skills v1.2 teaches from the other side. In this space the artefacts are markdown, so lock-in is nearly impossible to build, and the projects that leaned into portability grew fastest.

Rank #19 and rank #20

The repo one place above Hermes on the all-time list is mattpocock/skills at 204k stars. An agent runtime and a folder of markdown skills, side by side in the top 20 of all GitHub, both younger than two years. Models live on Hugging Face. The GitHub stars went to the two layers you customise an agent with: the runtime that runs it and the instructions that steer it.

I still owe this series a hands-on post; Hermes on a Windows machine is next on the list, and the survey’s conclusions deserve testing against a live install.