secondfry/claude-code-self-learn

self-learn is a plugin for Claude Code, the command line AI coding assistant, that adds a self-improving memory loop. It is a Python project with no GitHub stars yet, built by developer secondfry. The idea is based on an existing project called Hermes Agent, adapted with lessons from two other Claude Code plugins, claude-reflect and claude-self-improving-skills. Instead of learning during a live conversation, self-learn reads the transcript files that Claude Code already saves to disk after a session ends, and pulls useful information out of them once the session goes idle. Two kinds of information get extracted. Facts, meaning things worth remembering about a project or a user's preferences, are written into Claude Code's built in memory format, either to a global folder or to a folder specific to that project. Skills, meaning reusable techniques that worked well in a session, get written as SKILL.md files that Claude Code can reuse later. The extraction work happens on a schedule, roughly every thirty minutes, checked by a background dispatcher script. It waits until a transcript file has been untouched for a set number of minutes before treating the session as finished, so it never reads a conversation that is still in progress. It also tracks how much new activity has happened since the last check, so it only kicks off the extraction agents when there is enough new material. Installation is done from the command line using Claude Code's own plugin commands, and after installing, restarting a Claude Code session lets a background scheduler register itself automatically depending on the operating system: launchd on macOS, a systemd timer on Linux, or a Scheduled Task on Windows. The project requires Node and Python 3 to be available. The README highlights that no separate always-on server is needed and nothing gets added to the live conversation, since everything runs after the fact using files Claude Code already writes.