Answers

How team decisions reach the AIs doing the work

A decision reaches an AI when it is put in front of that AI before it acts — in the read the assistant makes at the start of a task, not in a document someone hopes it will find. Arroway keeps decisions as material of a project, marks the ones that must apply always, orders the rest by the task in front of the assistant, and says in the same breath how much fit and what did not.

Last updated November 14, 2026

A decision that was stored but never delivered changed nothing

Writing the rule down is the easy half. What matters is whether it was in front of the assistant at the moment it mattered — and most systems cannot tell you. Arroway's read comes with a count: how many memories came, how much room they took, how many did not fit. A rule that matched real work and was not delivered is treated as a defect of the product, measured and fixed, not as something a person has to work around by repeating themselves.

Always-on rules and everything else are handled differently

A rule whose breach is expensive or irreversible can be marked always-on: it appears in every read, in every session, whatever the task — which costs room every time, so it is a deliberate choice rather than the default. Everything else is ordered by the task the assistant declares, with the most recent as a tiebreaker. The daily log is ordered by subject too, so the history a new task should learn from is not pushed out by whoever wrote most recently.

Each project says what it is for, and that is what keeps decisions from leaking

Every project states its purpose in plain words, and a person can correct that statement. The assistant is told to write into the project the work belongs to — not into the nearest one it can find — and to read only from the projects the person belongs to. That is what stops a decision made in one context from governing work in another.

When a rule should be always-on, and when it should not

Ask what a breach costs. A rule that, once broken, sends money, deletes data or contacts a customer is a candidate: it should be in front of the assistant whatever the task is. A rule that only improves quality is not. Every always-on rule is paid for in every read by every task, including the ones it has nothing to do with, so a project with forty of them starts each task by spending most of its room before the task is even read. A working test: if you cannot name the specific harm that follows from missing it once, it is an ordinary rule and should be ordered by the task like the rest.

Three ways a stored decision still never arrives

The decision lives only in a chat transcript, so the next session has nothing to read. The decision is recorded as a story (“we talked about the CRM and felt it was too much”) with no statement of what is now true, so it never matches a task. Or it is recorded perfectly, then replaced in practice and never replaced in the record, so the assistant faithfully delivers the old one. The last is the most expensive, because it looks like the system working. A replacement is declared by pointing at the rule it replaces, never by writing a second version next to it.

What it looks like in practice

A hypothetical nine-person team decides on a Tuesday not to build a two-way sync with its customers' CRM, because the first two customers who asked said they only needed a weekly export. The reason goes into the record along with a condition that would reopen it. Six weeks later a new engineer asks an assistant to plan ‘the CRM integration’ for a prospect. Without the record, the assistant produces a competent plan for work the team already turned down. With it, the plan starts from the decision.

From a decision in a meeting to a decision in the next task

  1. State the decision and the reason in the same sentence — “no two-way CRM sync: the two customers who asked needed only a weekly export” — and say who made it. A decision recorded without its reason cannot be applied to a case slightly different from the one that produced it.
  2. Add the condition that would reopen it: a date, a number, an event. “Reopen if three customers ask in one quarter.” This is what stops the decision from asserting forever something that stopped being true.
  3. A person approves it, so it is filed as a decision and not as an assistant's guess. Until then it is a proposal and is served as one.
  4. Decide whether it is always-on. Almost never: a rule that appears in every read costs room in every read, so keep that for things whose breach is expensive or irreversible. This one is ordinary, and will be ordered by the task instead.
  5. Six weeks later the engineer asks for the integration plan. The assistant reads the project before it starts, declaring the task, and the decision ranks high because it matches the work in front of it. The assistant says so, quotes the reason and the reopening condition, and asks whether the condition has been met.
  6. Check the read, not the answer. It states how many memories it delivered, how much room they took and how many did not fit. If the decision matched the work and was not delivered, that is a defect in the product to report — not a reason to start repeating yourselves in every prompt.

Questions people ask about this

How do I know the AI actually received a decision?
The read says what it delivered and what it cut. And the audit screen shows when a memory matched real work after the last time it was delivered — which is the room running out, made visible instead of assumed.
What if two memories contradict each other?
The assistant is instructed to say so and bring both to a person, rather than pick one quietly. Contradictions also show up on the review screen, side by side, for someone to settle.
Does the assistant have to be told to read?
Once, while the habit is new: one sentence at the start of a conversation. After that the instruction travels with the product — the server repeats it in what it returns, and where your tool supports the plugin, the reminder comes on its own. It does not depend on each person's setup.
Install ArrowaySee how it works →

Where this is verifiable

The sanctioned product spec in the repository, the read protocol served by the MCP server, and the audit surface in the product, which reports what was delivered and what was cut.

https://www.arroway.app/en/answers/how-team-decisions-reach-ai

Continue exploring

See all answers