Proposal 2 of 3
A memory asking permission
The reading-chair footprint should stay under 34 inches wide.
Approve permits bounded recall. Reject keeps a review record. Expire blocks future use. Forget removes content and redacts its influence trail.
A human control layer for useful AI memory
Memory should ask before it speaks for you.
This example uses Unichat to turn inferred memory into a reviewable proposal. You decide the wording, purpose, exact scope, and expiry—then every future influence leaves a receipt.
StorageBrowser localStorage
Provider callsZero in this sample
Recall ruleApproved + exact scope + unexpired
Interactive browser sample
Review before recall.
Make a decision, reload the page, and the sample stays here. Choose the approved record to test expiry or forgetting; choose a pending proposal to approve or reject.
Awaiting review
Sample ready. Nothing leaves this browser.
The stewardship loop
Memory becomes a governed product behavior.
The sample separates inference from permission. A useful detail can be proposed without becoming recallable, and every later decision is explicit and reversible.
Show the whole claim.
Content, source, purpose, scope, and expiry arrive together for human review.
Consent is a state change.
Approve, revise, narrow, reject, or set a limit before recall becomes possible.
Influence leaves evidence.
A receipt identifies the approved memory used for a matching request.
Forgetting means redaction.
Content disappears while a safe tombstone blocks accidental resurrection.
Released August 19, 2026
Memory Steward browser sample
Memory Steward uses Unichat to make AI memory reviewable before it is useful: a concrete human surface for permission, bounded recall, transparent influence, and honest removal.
- Provider-free local sample with reload-safe state
- Approval, editing, scope narrowing, rejection, and expiry
- Same-scope recall with explicit cross-scope denial
- Influence disclosures and redacted deletion tombstones