nousresearch/hermes-agent is a self-improving AI agent repository built by Nous Research. It creates skills from experience, improves them while being used, searches past conversations, and develops its model of the user across sessions. The repository description is short: an agent that grows together with the user. The latest release in the record is v0.22.2. It is written in Python and follows the MIT license. GitHub stars are 255,671 and forks are 55,084.
What changed shows up in the release record. v0.22.0 The Threshold Release arrived on October 13 and carried a desktop experience that starts with a setup conversation on first run, connectors that run inside the app, local models that attach to the llama.cpp runtime with one click, and a plugin catalog of roughly 600 entries. The v0.22.1 patch that came out the next day filled in the Windows Store and the plugin catalog. The README says it supports multiple model providers, can be used from the CLI and on messaging platforms such as Telegram, Discord, Slack, WhatsApp, and Signal, and also supports scheduled automation, parallel sub-agents, and various terminal backends.
However, the given record does not say what v0.22.2 changed. The release entries with content stop at v0.22.0 and v0.22.1. Only the momentum evaluation item is filled in, at 9.9, while the novelty, tech, and quality items have no values. The 9.9 is an evaluation score, not a performance measurement. Nothing in this material shows what the self-improving feature actually changes in real conversations. Until the v0.22.2 release notes are checked, the sensible reading is that this is a self-improving agent project whose release flow keeps going.