Correctable understanding

Elephant Agent should become clearer over time, not more confident about stale claims.

Correctability is the product contract that keeps personal AI from becoming a hidden profiler. The user can remember, correct, forget, dispute, inspect, and ask why.

The correction moves

Move	Intent	Result
Remember	Carry a useful claim forward.	A new active Personal Model claim.
Correct	Replace a wrong or stale claim.	Old claim retires; corrected claim becomes active.
Forget	Stop using a claim.	Claim no longer shapes future replies.
Dispute	Mark uncertainty.	Elephant Agent avoids treating the claim as settled.
Explain why	Inspect support.	Source Episodes and Steps become visible.

Claim lifecycle

warning

A corrected claim should win over old recall. If current-turn recall finds a stale Step, that Step can explain history, but it should not override the active Personal Model.

No-match is a feature

no_match means Elephant Agent does not have reliable Personal Model support. That is often the right answer.

Search result	Good behavior
strong_match	Use the claim and keep the answer grounded.
weak_match	Avoid overclaiming; ask or qualify if needed.
no_match	Say the Personal Model does not contain support.

Foreground and background use the same path

Whether the update comes from chat, dashboard, init, or background learning, the important rule is the same: durable understanding changes through explicit Personal Model claim operations.

Where to correct

Surface	Best for
Chat / wake	Correcting naturally in conversation.
Dashboard Personal Model	Reviewing and correcting active claims.
Dashboard Curiosity	Answering or dismissing open questions.
Dashboard History / Why	Inspecting source support before changing a claim.