Hermes Agent completes the skill retirement cycle: use it and remember, if not used for 30 days then downgrade, archived after 90 days

According to Beating Monitoring, Nous Research’s open-source AI agent framework Hermes Agent has added a Curator feature, completing the final step in skill lifecycle management. Previously, one of Hermes Agent’s core selling points was that the agent could automatically create skill documentation after completing complex tasks, accumulating experience for later reuse—except skills only ever went in and never came out. Some users reported that after accumulating 146 self-built agent skills, the skill list in the system prompts for each conversation took up about 4,400 tokens; the more useless skills stacked up, the worse it got, with no elimination mechanism.

Curator handles this in two steps. First is a deterministic state machine that does not call a model: each self-built skill has counters such as usage count and the last call time. Skills that have not been used for 30 days are marked as stale, and skills that have not been used for 90 days are moved into an archive directory, from which they can be restored at any time. Second is model review: every 7 days, when the agent is idle, an auxiliary model is dispatched to scan the remaining skills, merging overlapping functions and patching outdated content, with up to 8 iterations. The review runs in a separate prompt cache and does not affect conversations users are currently having.

Curator only modifies self-built skills of an agent; it does not touch built-in skills or skills installed from the skill marketplace. It does not perform automatic deletion—at worst, skills are archived. Users can lock important skills with a pin command; once locked, the skills cannot be modified even by the agent’s own skill_manage tool. It is enabled by default and can be turned off. The PR is about 2,200 lines of code; all 56 tests passed. It was merged into main on April 29, and has not yet been released with the official version.