Hermes Agent: A Self-Improving AI Framework With Persistent Memory
Nous Research's open-source Hermes Agent framework combines persistent memory, reusable skills, and a real multi-agent architecture for self-improving AI.
Hermes Agent, an open-source framework from Nous Research, breaks from the typical stateless chatbot model by implementing a closed learning loop. The system combines reusable procedural "skills," cross-session persistent memory, FTS5-powered session search, and self-patching capabilities so the agent improves through actual use rather than resetting each conversation.
The framework ships with roughly 60 built-in tools — terminal access, file I/O, web search, code execution, image generation, and subagent delegation — and can run on a laptop, a cloud VPS, Docker, or serverless platforms like Modal and Daytona. A built-in cron scheduler lets agents run recurring jobs autonomously, such as daily reports or opportunity scans.
Hermes also supports multiple independent "profiles" — separate agents with their own memory, tools, and identity that communicate over HTTP using OpenAI-compatible endpoints — plus MCP (Model Context Protocol) integrations for connecting directly to external services like Feishu or email. For engineers, the key distinction is that this is a real multi-agent architecture running as independent processes on separate machines, not a single-prompt simulation.