komi-learn

komi-learn subcommand [options]

komi-learn is a Python CLI that adds a memory and self-improvement loop to coding agents such as Claude Code and Codex. It installs per-agent hooks that observe sessions and persist durable lessons across runs.The learning loop has four stages. Recall activates relevant lessons at the start of a session. Distill analyzes the session transcript after it ends and extracts durable lessons. Curate merges overlapping items and archives outdated ones. Share optionally publishes user-approved lessons to a GitHub-hosted, cryptographically signed community pool so they can be pulled by other users with `komi-learn sync`.Secrets, absolute paths, and one-off failures are filtered before content is stored or shared. `forget` archives a learning by default and only deletes it with `--hard`. `uninstall` removes hooks; `--purge` also removes stored learnings.

install

Set up the agent hooks. `--yes` accepts defaults. `--host` selects the agent (`claude`, `codex`).

Verify the installation and report problems.

Upgrade komi-learn and its hooks. `--check` previews only.

Display configuration and learning progress.

Adjust settings, either interactively or via `config set <key> <value>`.

Pull community-pool lessons.

Review pending shared contributions.

Archive a learned item. `--hard` deletes it permanently.

Remove hooks. `--purge` deletes all stored data.

komi-learn modifies the host agent's configuration during `install`. Run `komi-learn doctor` after an upgrade to confirm hooks are still valid. The community pool is opt-in, but be aware that lessons you queue for sharing may include excerpts of code or commands from your sessions; review the queue before syncing.

claude, aislop