v0.5.3

• Alfred can now keep the most useful lessons on its own. A second model reads each candidate lesson, and when it is confident the lesson is sound and worth keeping, the lesson is saved into recall without waiting for you. Anything it is unsure about still waits in the review queue, and every automatic save can be undone.
• The lessons that change how Alfred works are the ones it now saves for itself. When a lesson would actually change how the agents behave next time, and the reviewer is confident it is safe, Alfred saves it straight away rather than letting it sit unreviewed. Weak or one-off lessons are still held back.
• The memory server is more reliable. Alfred now uses the local memory server purely to store and look up lessons, and no longer lets the small on-device model rewrite or merge them. The lessons you and the fleet trust are kept exactly as written.
• A full disk no longer wedges the fleet. When space runs critically low, the emergency cleanup now also clears regenerable build and download caches across the machine, so the agents can keep working instead of getting stuck.

Full changelog: CHANGELOG.md