Hermes Agent: Self-Improving Agent with Closed-Loop Learning
Source: https://github.com/NousResearch/hermes-agent Author: Nous Research Date: 2025-2026
Summary
Hermes is Nous Research’s self-improving agent with a closed-loop learning system. It creates skills from experience, improves them during use, and persists knowledge. Compatible with any LLM provider. Supports multiple interfaces (terminal, Telegram, Discord, Slack, etc.) and six execution backends.
Key Claims
- Only agent with built-in learning loop: skills created from experience, improved during use
- Persistent user profiles and procedural memory across sessions
- Six terminal backends: local, Docker, SSH, Daytona, Singularity, Modal
- Parallel workstreams via isolated subagents
- Built-in cron scheduling for unattended tasks
- Batch trajectory generation + RL environment compatibility for tool-calling model training
- Compatible with OpenAI, Anthropic, OpenRouter
Connection to Other Sources
Represents implementation of Agent Memory concepts. Compare to RL-trained memory agent and Mastra observational memory.
Concepts
- Agent Memory — persistent skills and user profiles
- Coding Agents — agent with multiple execution backends
- RL Infrastructure — RL environment support for training