asimons81/hermes-dreaming

Hermes Dreaming is a tool that lets an AI system propose updates to its own memory, stored facts, user preferences, and skills, but holds those proposals in a waiting area for a human to review before anything actually changes. Think of it as a suggestion box with a lock: the AI can write ideas for how to update itself, but nothing gets applied until you explicitly say so. The core loop has three steps. First, you run a "create" or "review" command that points the tool at some source material, such as recent conversation logs or notes. The tool reads those sources and drafts a set of proposed changes, saving them to an artifact directory as plain files you can inspect. Second, you open those files and go through the proposals: approve the ones that look right, reject the ones that do not, and leave notes about why. Third, you run an "apply" command, which writes only the approved changes to the live memory state and backs up what it replaced. This staged approach is the point of the project. AI memory systems that update themselves automatically carry the risk of slowly drifting in unintended directions. By requiring an explicit review and apply step, this tool keeps a human in the loop for every change. Rejected proposals are archived, not deleted, so you have a full audit trail of what was considered and why it was turned down. The tool ships as both a standalone Python command-line program and as a plugin for Hermes, an AI memory management system. A cron job installer lets you schedule nightly dry-run reviews that stage proposals automatically but never apply them without your approval. An offline quickstart demo shows the full flow without requiring an API key or an external AI model. The project is open source under a standard license and includes safety documentation explaining what the tool can and cannot modify.