Hermes: self‑improving agent pitch

Hermes is being pitched as an artificial intelligence agent that does not just answer prompts once, but stores what worked and reuses it later. (youtube.com) Nous Research’s GitHub repository describes Hermes Agent as “the agent that grows with you,” and says it can create skills from experience, search past conversations, and build a user model across sessions. The repository showed about 82,500 stars when it was crawled on April 15. (github.com) The April 15 YouTube video and a same-day Geeky Gadgets write-up both framed that learning loop as the product’s main feature. Geeky Gadgets said Hermes uses a system called Generalized Action and Prompt Adaptation, or GAPA, to review and refine actions every 15 calls without changing the base model. (youtube.com) (geeky-gadgets.com) That pitch lands in a crowded “agent” market where many tools can call software, browse files, and run commands, but often lose context when a session ends. Hermes is being sold on the idea that memory and repeated use, not just bigger models, can make an assistant more useful over time. (github.com) (hermes-agent.nousresearch.com) In plain terms, Hermes is trying to turn one-off instructions into reusable habits. If a user asks for the same workflow again and again, the system says it can package that pattern into a skill so the next run needs less hand-holding. (geeky-gadgets.com) (hermes-agent.nousresearch.com) The memory piece matters because it is the difference between a chatbot and a persistent assistant. Hermes’ public materials say it can keep conversation history, search prior sessions with full-text search, and store user data locally depending on setup. (youtube.com) (github.com) The project is also being marketed as flexible rather than tied to one model vendor. Nous Research’s documentation says Hermes can be installed with a one-line installer, while Geeky Gadgets says it supports providers including OpenAI and Anthropic and requires Windows Subsystem for Linux 2 on Windows for the recommended setup. (hermes-agent.nousresearch.com) (geeky-gadgets.com) The pace of updates has been fast. GitHub shows Hermes Agent v0.7.0 released on April 3, 2026, and v0.8.0 released on April 8, 2026, with additions including model switching, background task notifications, and approval buttons for Slack and Telegram. (github.com 1) (github.com 2) The open question is whether “self-improving” means dependable gains in real work, not just more automation layers. For now, the public evidence is mostly Nous Research’s own documentation, GitHub materials, and enthusiastic secondary coverage rather than independent benchmarks. (github.com) (geeky-gadgets.com)