ConceptAvailable
Memory Health
Trust signals travel with every retrieval. Memory health verdicts are part of every read so downstream tools see the same state as the in-app UI.
The ten dimensions
- Staleness, old enough that the world likely changed
- Conflict, multiple memories disagree
- Low confidence, extraction confidence below the band
- Missing source, a claim with no source link
- Duplicate, two rows describe the same item
- Archived, the row is preserved but not active
- Cost-capped, embedding budget reached for the workspace
- Pending, a candidate not yet approved
- Reviewer-flagged, explicit reviewer warning
- Source-status, the underlying source is broken or removed
How it surfaces
Verdicts ride on the same response envelope as every other read. Context packs downweight unhealthy items and emit explicit warnings.
For runnable code samples and the developer reference, see /developers. For the trust model in depth, see /security.