Recall

Quality

Agent evaluation

These scenarios run live against the real memory pipeline — real embeddings, real database writes, in a throwaway isolated session that is cleaned up afterwards. Nothing here touches your own memories.

Relevant memory retrieval

A stored budget preference is retrieved for a product request.

not run

Irrelevant memory rejection

An unrelated factual question retrieves nothing.

not run

Duplicate memory prevention

Re-stating a known fact merges instead of inserting.

not run

Memory update

Changing a value edits the existing record in place.

not run

Memory deletion

Deleting removes the row from the database.

not run

Cross-user isolation

Another session's memory is never visible.

not run

No-memory fallback

With nothing stored, retrieval returns empty and the agent still answers.

not run