Review inbox: approve proposed notes
An AI or agent can propose a new note rather than writing it — the proposal lands in a review inbox where you approve or reject it. This is the write side of AI you can trust: the AI proposes, the human approves, and the approval is signed.
The inbox is a channel, not a wall. It covers proposed notes. It does not intercept every write a model might make. The real protection is the access level you grant: your AI is read-only by default, and while it stays read-only nothing it does can touch your vault at all. If you grant read-and-write, the other writing tools — writing to the graph, running an ingest, and the rest — take effect immediately, with no inbox stop. Grant that level only to a client you trust, and see What your AI sees for where the switch is.
Why proposals, not direct writes
Letting a model write straight into your knowledge base is how a second brain fills up with plausible-but-wrong notes, silent duplicates, and edits you never agreed to. For new notes — the highest-volume, lowest-accountability kind of AI write — Altretta puts a human checkpoint between "the AI generated something" and "it's part of my vault." The generation is cheap; the acceptance is the accountable act, so acceptance is what gets signed.
How a proposal flows
The AI proposes
An in-app assistant, an agent, or a connected MCP client submits a note it wants to add. It lands in the review inbox as a proposal, staged outside your indexed notes rather than added to your vault.
You see a badge count
The inbox shows a badge with the number of pending proposals waiting for your decision, so nothing queues up unseen.
You review with context
Each proposal shows similar existing notes alongside it, so you can spot overlap before you accept — a built-in dedupe aid against near-duplicates.
You approve or reject
Approve, and the note is added to your vault and the approval is recorded as a signed action in the DAG. Reject, and the proposal is archived — never silently deleted.
Similar-notes dedupe
A brand-new note that restates something you already wrote is one of the most common ways knowledge bases rot. For every proposal, Altretta surfaces the similar existing notes it found in your vault. Before approving, you can see at a glance whether the AI is adding genuinely new knowledge or duplicating a note you'd forgotten — and either merge the idea into the existing note or accept the new one deliberately.
Approve, reject, and the audit trail
Approval is a one-way commitment into your signed history — treat it as "I stand behind this." Rejection is reversible and non-destructive: a rejected proposal is archived, so you can revisit it, not lose it.
Every decision is retained as an approval audit trail: what was proposed, what you approved or rejected, and when. Approving also stamps the note itself with its approved status and the time it was reviewed, and records the approval as a signed action so it inherits the guarantees of the rest of your signed history — hash-linked, tamper-evident, and attributable to your identity.
Signing the approval is best-effort. If the signed record can't be written at that moment the note is still approved, but it will show as unsigned in its history rather than pretending to a signature it doesn't have.
The trust workflow, end to end
The review inbox is one part of a wider write-side story:
- Read side — What your AI sees shows the context a model reads and lets you exclude folders from it, and the activity log records what it queried.
- Write side — the access level decides whether a connected AI can write at all; the review inbox gives proposed notes an explicit human approval on top.
If a proposal seems to be missing similar notes it should have matched, rebuild your vault index so the dedupe check is working from your current knowledge.
Related
AI you can trust
Back to the AI section overview.
What your AI sees
The read-side transparency panel — the exact context behind an answer.
Activity log & consistency review
A running record of what the AI queried, plus contradiction detection.
Signed history (the DAG)
Where each approval is recorded as a tamper-evident signed action.
Glossary
Definitions for proposals, actions, the DAG, and more.