Memory review
Memory review is where Asterel’s companion promise becomes inspectable. The goal is not to expose every raw event by default. The goal is to let an operator see why a remembered fact exists, correct it when wrong, and prevent private context from leaking into the wrong surface.
Review model
Section titled “Review model”Memory should be read in layers:
compact view -> timeline / provenance -> full evidence recoveryThe compact view is enough for most live operation. Provenance and timeline help answer “why does the companion believe this?” Full evidence recovery is for review, correction, deletion, or incident analysis.
Correct, forget, delete
Section titled “Correct, forget, delete”Use different operations for different intent:
| Intent | Operator meaning |
|---|---|
| Correct | The old fact was wrong or stale; keep lineage and make the current view right |
| Forget | The fact should not be used again, subject to backend and policy semantics |
| Delete / hard removal | Stronger removal path when supported and policy permits it |
Do not silently overwrite facts that shaped prior behavior. Correction lineage is part of the trust story.
Public/private exposure
Section titled “Public/private exposure”The companion may use private memory as latent grounding in a private context, but that does not mean it may say the fact in a public channel. Exposure review should ask:
- where did this memory originate?
- is the current turn public, thread, DM, or gateway?
- was the fact corrected or marked sensitive?
- did response finalization block or revise the draft?
Self-amendment review
Section titled “Self-amendment review”Sometimes the useful memory is not a fact about the user. It is a lesson about how the companion should approach this person or surface next time. Those lessons are self-amendment candidates.
They should stay governed:
- generated as reviewable candidates;
- bounded and redacted rather than copied raw transcripts;
- approved by an operator before durable persistence;
- stored as private procedural memory, not as character-core mutation.
Do not turn a user correction into an automatic intimacy upgrade. Repair should improve future behavior while preserving the user’s freedom to correct, reset, leave, or ask for forgetting.