
Gives Claude a SQLite backed memory across sessions so it stops suggesting tools you've already rejected or repeating mistakes from last week. Exposes tools to record decisions, store lessons with verification status (claim/evidence/confirmed), and run preflight pattern matching against your task before Claude starts work. The verification gating is the interesting bit: unverified guesses weigh 5x less than confirmed lessons, and write operations re-read from the database before returning success to catch silent storage failures. Includes stale memory detection that warns when your daily markdown logs are newer than the last Amplifier write. Local only, no cloud sync. Useful if you're working with Claude on multi-day projects where context actually matters between sessions.