Mastra 1.66.0 adds redaction hooks for agent memory

Mastra released version 1.66.0 of its memory package with hooks that redact messages before they reach the model or get saved. Anyone running agents with PII in memory has been asking for this, since the only fix before now was forking the library.

The new beforeObservation and beforeReflection hooks run before a message reaches the model or gets written to storage. AfterObservation and afterReflection run after, for reshaping what's already been recorded.

That ordering matters. Filtering sensitive data out of memory once it's already stored doesn't undo the exposure. A hook that runs before persistence stops the data from being written in the first place.

The same release adds idempotent batch deletion methods, deleteFeedback() and deleteScores(), with optional scoping by organization or resource, working across ClickHouse, DuckDB, and Postgres. Trace queries also gained same-span predicates on model, provider, duration, outcome, identity, and lineage, per the release notes.

Together, the hooks and the batch deletion cover what a compliance request needs: stop new sensitive data from arriving, and remove what's already there. A retry on a failed delete call shouldn't throw or double-count. Teams usually find that out the hard way, after a partial failure already happened.

The release notes don't say whether the hooks apply to messages already stored before the upgrade. Teams with existing PII exposure still need the batch delete path to clean up what's already there.