Answers
Human-governed memory for AI work
Governed memory means a person decides what counts as a rule, and the AI cannot promote its own inference to that status. In Arroway an assistant writes freely, but what it writes on its own arrives as a proposal; a human turns proposals into standing rules — or corrects them — in a review that takes minutes and happens after the fact, never as a gate the work waits on.
Proposal and decision are different objects
When an assistant records something it inferred, the entry is a proposal. When the person stated or sanctioned it in the conversation, it is a decision. The distinction is stored, shown, and served differently to the AI, so an assistant reading later cannot present a proposal as something the team decided.
The human is a veto, not a bottleneck
Review is a short pass over what the system already did by itself: approve, correct, or refuse per line, with the standing memory and the proposed replacement side by side. A design that required approval before memory worked would make the account owner the bottleneck, and it collapses the moment several people are writing.
Every rule carries what would kill it
A memory is stored with the condition that would end it or force a re-check — an explicit revocation, a named trigger, a date. Dated status can be marked to stop being served once it expires, so nothing asserts as current what stopped being true. Provenance sits next to the rule for the person auditing it, and never travels into the AI's read, where it would compete with the rule itself.
What it looks like in practice
An assistant infers a naming convention from the code and records it. It shows up in your review as a proposal; you rewrite one sentence and sanction it. From then on it governs, and every assistant on the project reads the version you approved.
Questions people ask about this
- Can the AI change a rule I approved?
- It can propose a replacement, pointing at the memory it would replace. The standing rule keeps governing until a person accepts the replacement.
- Do I have to review everything for the memory to work?
- No. The record works immediately; review is how you cut what should not have entered. Curation is the product's job, and manual review is not meant to become recurring work for the owner.
- What stops the memory from filling with near-duplicates?
- Writing is checked at the door — related memories are shown before a new one is saved, and a replacement is declared by reference rather than guessed from titles. Automatic auditing of the collection is a product responsibility, and detectors enter it only with recurrence measured in production.
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