Measures

A revenue figure answered faster is worth nothing if the number is wrong.

Jargon holds the words your business uses in its own way, and it is deliberately soft about them: each note is a pointer to what a word is getting at, never a closed definition, because a short note on a broad term read as the whole of it has an assistant correcting the people whose word it is.

That is right for a word and wrong for a number. Nobody is harmed by a loose account of what "onboarding" covers. Somebody is harmed by a loose account of active merchant, because an answer that says 412 when the finance pack says 388 is not vague — it is wrong, and it is wrong in the way that costs a decision. Worse: nothing in the answer tells you which it is.

So there is a second register for the other kind of vocabulary — the words that resolve to a quantity.

What a measure is

A name, what it counts in business terms, and the rule: the filter, the exclusions, the grain and the unit that make it your figure rather than the obvious one.

Name	Active merchant
Counts	Merchants who took a payment in the period
Rule	Distinct merchants with at least one settled payment in the calendar month
Unit · grain	count · per month
Excludes	our own test accounts · internal transfers
Expression	select count(distinct merchant_id) from settlements where … — optional, against a connected database

The rule is in words, so it outlives the database it was written against. The expression is optional and does something the words cannot: it makes the rule executable, so the assistant runs your query rather than writing its own.

Settled, or not yet

A measure is settled when somebody stands behind the rule, and a proposal when nothing does yet. Only a settled rule reaches the assistant; a proposal is offered with the fact that nobody has confirmed it said out loud.

What decides which is where the rule came from, not a box anybody ticks:

• A person states it — on the Measures page, or in a conversation with the assistant. That settles it, where they say it. Somebody explaining a figure in plain words is that person standing behind it, and making them then confirm the thing they have just said only teaches them that saying it did not count.
• Two independent sources agree — a rule a second source restates on its own becomes your definition with nobody asked. Two systems agreeing is the strongest signal a memory system can have, and sources are named rather than counted, so a nightly pass re-reading one table is not two systems agreeing.
• A machine inferred it — read off a column name, lifted out of a document, or written by a scheduled run with nobody watching. That waits. A rule inferred from the shape of your data is a guess about the most consequential thing in the register, and a guess presented as your definition is worse than an empty register, because every answer afterwards is confidently wrong in the same way.

It is the same test Memtro applies to a remembered fact: a person was there and the request had read nothing external, so the conclusion is as good as they are. Anything else produces a claim, and a claim waits.

When your systems disagree

The expensive failure in a company's numbers is not an undefined figure — that at least announces itself. It is two definitions of one name, each defensible, living in different systems, with nobody aware there are two. The CRM counts a merchant active on a flag somebody set by hand; the ledger counts one that transacted in the last thirty days. Both are right.

A rule that disagrees with a settled one does not overwrite it. Both are kept, the figure is marked as having two definitions in use, and the assistant then gives the figure each way with its source rather than picking one. Settling it is a person's decision, and the rule that loses stays on the entry — because "why does the board deck say something different?" is a question somebody will ask.

Two things keep that queue honest:

• A rewording is not a disagreement. "Settled payments less refunds" and "settled payments, net of refunds" are one rule. Raising that as a dispute puts a decision in front of somebody that does not exist, and teaches them to clear the queue unread — which costs the real disputes their only claim on anybody's attention.
• You can revise your own rule; you cannot quietly overwrite a colleague's. Same person, same figure: a correction. Anybody else: both rules stand.

Where they come from

You do not start with an empty page. Memtro proposes the figures your own systems already hold, two ways:

1. From what people say. The pass that reads every exchange for durable facts also reads it for a rule somebody has just stated. Say "active means they took a payment in the last 30 days, excluding our test accounts" in a conversation and the figure is recorded and settled, because you said it — nobody has to ask for it, and the assistant does not have to think to reach for a tool mid-answer.
2. From use. Memory records the amounts, dates and counts your conversations and connectors mention, against the entities they belong to. A numeric key that turns up on many entities from more than one system is, by definition, a figure your company keeps and quotes — and the fact that two systems hold it is exactly the condition under which the two ways of computing it differ and nobody has noticed.
3. From your schema. On its weekly pass, Memtro reads the catalogue of each connected Postgres, MySQL or SQL Server database — the table and column names, not the data — and proposes the measures a business running those tables plainly has: a settled_amount on transactions, a status on merchants. This is the only route that finds a figure before anybody asks for it, and the only one that can offer you an expression.

Only the first of those settles anything, and only because a person said it. The other two fill the queue: you say which ones are right, or a second source does.

How the assistant uses them

Every settled measure contributes one line to the assistant's prompt — its name, what it counts, its unit and grain. The rule is not in the prompt: the assistant reads it with memtro_measures when a question actually turns on that figure.

The instruction that matters is the one the register exists for: read the rule before you give a number, and compute it that way. A model asked for your figure will otherwise compute one from the obvious reading of the words and present it with exactly the confidence of the right answer. Where a measure carries an expression, the assistant runs that rather than inventing a query.

That instruction is also served in the MCP handshake, because a coding agent connecting over MCP never sees Memtro's prompt — and that is precisely the surface where it would otherwise write its own SQL and hand back a number nobody downstream can check.

Every answer says what it could rest on

Separate from the register and running whether or not you keep one: each answer is graded on what the question named and what memory actually has behind it. The names and identifiers you used, checked against what retrieval found; the figures the question turned on, checked against the register.

What is uncovered is said rather than worked around — "memory holds nothing about Fabrikam, so say that rather than answering around it" — and the grade and the uncovered names come back on the response as memtro.grounding, for a client that wants to show them. It costs one pass over text already in hand: no extra model call.

It is a measure of coverage, never of correctness. A full grade means every subject the question named has something behind it, not that what is behind it is right. A number presented as a confidence would be exactly the kind of invention this page is about.

Asking is how it fills

A question that turns on a figure nothing defines leaves a proposal with no rule and a count of one. The next person asking makes it two. "Fourteen people have asked us for net revenue and nobody has ever said what it means" stops being an impression somebody has and becomes a row on a page, ordered by how often — each one an answer somebody has already been given with the arithmetic guessed.

Rule reads are counted separately, at the moment a rule is handed over, so the register can tell you which of your definitions are load-bearing and which nobody has needed since the day they were written.

A settled rule that stops working

A rule can be settled, in date, and uncomputable — the day somebody renames the column it sums. Nothing about that day looks different: no error and no empty answer, because the expression is not run until a question needs it, and then it fails inside one step the assistant works around.

So the weekly schema pass, which already has your catalogue in hand, checks that the names each expression uses still exist. What is missing is named on the entry and marked in the assistant's prompt — which matters, because the prompt is what tells it to prefer your expression over its own.

Keeping them current

Settling a rule asks how long before somebody should look at it again. Past that it is marked as needing a check, and the assistant says so when it uses the figure. Correcting a rule clears the last check automatically: what anybody verified was a different measure.

A measure belongs to the whole organisation, or to one team where only that team counts that way. A team's measures are invisible to everyone else.

It is already running

There is nothing to switch on and no page to visit. Measures is not a feature you evaluate: along with your vocabulary and your sequences it is part of what memory holds — the half of what your organisation knows that nobody writes down. None of the three is on the Features page or in the sidebar, because none of them is a thing anybody administers.

Two reasons they are not switches. Each costs nothing until there is something to find: proposing a figure needs the same number recorded against four entities from two separate systems, which is one database query that spends nothing when it comes back empty. And being off is not neutral the way the other registers' being off is — an organisation without Commitments simply has no promise register, whereas without this one you get numbers computed the obvious way, with the confidence of the right ones, in the part of the product a board paper is written from. That is not a thing to discover by reading a settings page.

The pages are still there — /dashboard/measures, /dashboard/jargon, /dashboard/procedures — as the audit surface, where you can read the whole register, settle a disagreement and see the figures people keep asking for. They are reached by a link rather than by browsing, the way the Connections page is after an OAuth redirect.

Reading the register is every member's; Certify measures is a permission on its own, and the only one in this group that is off by default — a settled rule is what every later answer computes from, so it belongs with whoever owns the number (Teams and permissions). The tools are in the tools reference.