Working Memory Review Checklist

System or chapter:
Owner:
Reviewer:
Date:

Fit check
[ ] Task spans multiple turns, tools, retries, approvals, or workflow steps
[ ] Goal has observable success criteria
[ ] Runtime can inspect and replay state
[ ] State can be stored safely
[ ] Durable memory promotion is separate from working memory

Goal and state
[ ] Goal ID recorded
[ ] Run ID recorded
[ ] Owner recorded
[ ] Success criteria recorded
[ ] Constraints recorded
[ ] Risk class recorded
[ ] Current step recorded
[ ] Completed and pending steps recorded
[ ] Open questions recorded
[ ] Evidence references recorded
[ ] Tool result references recorded
[ ] Approval references recorded
[ ] Budget state recorded
[ ] Stop reason recorded

Events and replay
[ ] State changes through typed events
[ ] Events are idempotent
[ ] Tool results create evidence events
[ ] Approval, denial, and expiry create events
[ ] Cancellation creates a terminal event
[ ] Replay reconstructs current state
[ ] Resume does not repeat unsafe work

Compression and retention
[ ] Working memory is smaller than transcript
[ ] Required fields survive compression
[ ] Raw payloads are referenced rather than copied when possible
[ ] Sensitive data is redacted or avoided
[ ] Stale state loses to fresh evidence
[ ] Completed run state is disposed, archived, or reviewed by policy

Promotion
[ ] Promotion to durable memory requires explicit policy
[ ] Source trust is checked before promotion
[ ] Consent or tenant policy is checked before promotion
[ ] Retention, correction, and deletion paths exist

Evaluation
[ ] Fixture covers multi-step goal creation
[ ] Fixture covers tool-result state event
[ ] Fixture covers retry idempotency
[ ] Fixture covers resume after interruption
[ ] Fixture covers approval wait
[ ] Fixture covers budget exhaustion
[ ] Fixture covers stale state
[ ] Fixture covers no automatic durable-memory promotion

Release decision
[ ] Green: ready for controlled use
[ ] Yellow: limited use; improve replay, compression, or promotion policy
[ ] Red: demo only

Required changes:

Reviewer notes: