Candidate Review (Ember)
Ember is the review capability in the memory system. It sits between raw capture and canonical memory to keep quality high.
Why this capability matters
- Prevents raw notes from polluting long-term memory
- Gives teams a clear promotion workflow
- Creates shared ownership of what becomes reusable guidance
What Ember does
- Surfaces candidate observations for review
- Supports accept, reject, or refine decisions
- Preserves links to source observations and sessions
- Requires explicit human decision before promotion
Operating contract
- Inputs: Candidate observations promoted from capture
- Decision point: Accept, reject, or refine promotion candidates
- Outputs: Reviewed candidates with decision, rationale, and confidence
- Evidence: Review trail that links candidate outcomes to source observations
Review criteria teams care about
Use these checks before promotion:
- Is this reusable across multiple tasks or engineers?
- Is it still correct under current architecture?
- Does it include enough context to avoid misuse?
- Is it already captured elsewhere in canonical memory?
Example promotion decision
Candidate: "Retries hide payment provider 429 limits"
Decision: Promote
Reason: Repeated production issue across two services
Confidence: High
Owner: Platform team
Positioning note
Externally, this is your quality gate for memory. Internally, Ember is the component name that implements the workflow.