Answers
AI memory a person approves before it counts
Human-sanctioned memory means a person decides what becomes a rule, and the AI cannot promote its own guess to that status. In Arroway an assistant writes freely as it works, but anything it worked out on its own is filed as a proposal. A person turns proposals into standing rules — or edits them, or throws them out — on a review screen, after the fact. The work never waits on that review.
Last updated November 14, 2026
A proposal and a decision are two different things, and they stay different
When an assistant records something it inferred from the code or the conversation, it is filed as a proposal. When you said it, or approved it, it is filed as a decision. The two are stored apart, shown apart on screen, and handed to the next AI differently — so an assistant reading the project six months from now cannot pass a guess off as something the team agreed.
Review is a veto, not a gate
Nothing waits for you. Your AI reads and writes while you sleep; review is where you go afterwards to correct what should not have gone in. The screen gathers what is worth a look — what came in this week, proposals nobody confirmed, entries that have reached their expiry date, and what no AI has read in a month. You approve, edit or delete line by line, and a replacement is shown next to the rule it would replace. Skip a week and nothing breaks: meanwhile Arroway keeps the rules that matter most in front of your AI.
Every rule carries the thing that would kill it
A rule is written together with the condition that ends it or forces a re-check — a date, a named trigger, an explicit revocation. When that comes due, the rule goes back to you instead of asserting forever that something is still true. Where an entry came from and who approved it sit next to the rule for whoever is auditing; that never travels into the AI's read, where it would compete with the rule itself.
When review is the wrong tool, and when it has stopped meaning anything
Review is for what ought to govern future work. State that only matters for today — which file you were halfway through, a half-finished job — belongs in a handoff, not in the rules, and secrets belong nowhere in the memory at all. Review also stops meaning something the moment it becomes a bulk button: if you approve forty proposals in a minute, ‘sanctioned’ is just a label on things nobody read, and the assistant six months from now will trust them as if you had. A short, regular pass of line-by-line decisions is worth more than a long, rare one. The memory is built so that skipping a fortnight costs nothing, which is exactly why it must not be turned into a rubber stamp.
A common error: approving what you have not read
The proposal sounds plausible and the title matches something you believe, so you approve it. Then the body contains a detail that is wrong — a version number, an exception, a boundary — and now it is a decision with your authority behind it. The defence is not suspicion of the assistant; it is that approval is a person's act of reading. If you do not have time to read a proposal, leave it a proposal. It costs nothing to wait, and it keeps its status honest.
What it looks like in practice
A hypothetical engineering team of seven has an assistant working in a repository with thirty services. While adding the thirty-first, it notices that every service is named with the team prefix and a noun in the singular, and it records the convention. The record appears in the next review marked as a proposal: an assistant worked it out, and nobody agreed to it. That distinction is the whole page, and the steps below show what happens to it.
From a guess to a rule, in the order it happens
- The assistant records the convention while it works, without waiting for anyone. It files it as a proposal because it inferred it from the code and no person said or approved it.
- The proposal turns up in the review with the other things worth a look: what came in this week, proposals nobody has confirmed, entries near their expiry date and what no assistant has read in a month.
- A person reads it. The convention has an exception the code does not show: the two services that wrap an external provider keep the provider's name. A proposal that is almost right is edited, not accepted or rejected.
- They approve their own version. From now on it is a decision, stored apart from the proposals and handed to the next assistant differently.
- They add the condition that would reopen it, for example “when the monorepo migration lands”. That is the date on which the rule returns to a person instead of asserting forever that it is true.
- Months later a different assistant starts a task in the repository, reads the project and follows the person's version. The original guess is still on the record, marked as what it was, for whoever audits it.
Questions people ask about this
- Can the AI change a rule I approved?
- It can propose a replacement, and it has to point at the rule it would replace instead of quietly writing a second version. Yours keeps governing until you accept the swap. And if it proposed something and you told it it was wrong, it can withdraw its own proposal — the withdrawal stays on the record, because “it suggested this and backed off” is worth knowing.
- Do I have to review everything for the memory to work?
- No. Recording works from the first day whether or not you ever open the review screen; meanwhile Arroway decides on its own which rules go in front of your AI. Review is how you cut what should not have entered, and it is a list Arroway assembles for you rather than a queue you have to clear — you can leave it a couple of weeks.
- What stops the collection from filling up with near-duplicates?
- Two things, and neither of them is you noticing. Before a new entry is saved, the related ones are shown to whoever is writing, so a second version of a rule that already exists has to be a deliberate act; and a replacement is declared by pointing at what it replaces, never guessed from similar titles. Whatever gets past that lands on the review screen as a collision for a person to settle. Nothing is merged or deleted automatically.
Where this is verifiable
The sanctioned product spec and the curation decisions recorded in the repository, plus the review surface in the product itself. The behaviour described here is what the tools enforce, not a roadmap.
https://www.arroway.app/en/answers/human-sanctioned-ai-memory