Contradictions

Every company believes two things that cannot both be true. Not because anyone was careless — because the two halves arrived months apart, from different systems, and nobody has ever read them side by side. The CRM says forty-one merchants are travel agents; the website says you serve tour operators. A renewal is March in the contract and June in the deal record. Each is defensible on its own. Together they are a question somebody should answer.

Memtro already holds both halves in one memory, which makes it the only place they can be compared. Dashboard → Memory → Contradictions is where it asks.

Contradictions are off until you turn them on, under Organisation → Features.

What it looks for

Consolidation — the nightly pass that keeps memory tidy — folds together statements that say the same thing. This is its mirror image: statements close enough to be about one subject, but saying different things.

Each night the pass looks out from the newest memories to their nearest neighbours in meaning and keeps the pairs in a band. Closer than that band and the two are one thing said twice, which consolidation handles. Further away and they are about different things, so saying different things is no surprise. Pairs about two different entities are dropped outright: "Acme has three open deals" and "Baltic has seven" differ in every figure and disagree about nothing.

What survives is ranked before anything is spent on it. A pair whose figures, years, quantifiers or denials differ is a better use of a model's attention than two statements that merely sit near each other, and a pair drawn from two different systems is better still — a CRM record against a scraped page is where a business actually contradicts itself, while two rows of the same sync are usually just two rows.

A model then judges the shortlist, and is told to be conservative. These are not contradictions, and it is instructed to say so: statements about different parts of the business, one more specific than the other, a change over time where one plainly describes an earlier state, two ways of counting that could both be right. A register full of ordinary variations is one people learn to scroll past.

What a row says

The question, as a colleague would ask it, and then both statements — each with the system it came from and the day it was recorded. That provenance is not decoration: choosing between two statements is impossible without knowing who said what, and when.

Each row carries how much it matters. High means somebody could act wrongly on it today — a customer-facing claim, a price, a contract term, a figure that gets reported. Medium is an internal fact people rely on. Low is stale wording nobody acts on.

Answering one

Four ways out, and they do different things to memory.

• A is right or B is right supersedes the other statement. This is the point of the whole feature: a system told which of two beliefs is wrong, and keeping both, has learned nothing. The losing statement is superseded rather than deleted, so what did we believe in March? still answers correctly.
• Both are true leaves both standing — the common answer when they describe different periods, or different parts of the business.
• Not a contradiction closes it and leaves memory alone, which is how the register stays trustworthy. A question nobody can dismiss is one people learn to ignore.

Whatever you choose, a note saying what the answer rests on is kept with it, along with who answered and when. A pair raised once is never raised again — answered, dismissed or otherwise — so the list only ever grows with things nobody has looked at. Ask again reopens one whose answer turned out to be wrong.

There is also Investigate, which hands the whole thing to the agent: both statements, both sources, and an instruction to work out which is right from the systems themselves and say which it checked.

The four answers, and what each does to memory

Answer	What happens
A is right	B is superseded: kept as history, stamped with the day it stopped being current.
B is right	The same, the other way round.
Both are true	Both stand. The right answer when they describe different periods, or different parts of the business.
Not a contradiction	The pair is closed and memory is left alone.

Only the first two change what the company believes, and they are the point of the feature: a system told which of two beliefs is wrong, and keeping both, has learned nothing.

From chat, the API and MCP

Three tools, available wherever Memtro answers:

• memtro_contradictions lists the pairs with both claims and their provenance. It says plainly that a pair is two remembered statements rather than a finding, and that the system which would settle it should be checked first.
• memtro_answer_contradiction settles one, and insists on a note saying what the answer rests on.
• memtro_raise_contradiction records one the agent finds while doing something else. Where it can pass the memory ids behind both statements, answering can supersede the loser; where it cannot, it says so rather than pretending otherwise.

An empty register is reported as a fact about the nightly pass, not about your business. Asked for contradictions when none are recorded, Memtro says what the register holds and offers to compare two systems — it does not answer that there are none.

What it is for

The obvious use is hygiene: memory that corrects itself is memory you can keep trusting. The less obvious one is that most of these pairs are not memory problems at all. A website that no longer describes the business, a renewal date two teams hold differently, a support promise the numbers do not support — those are the business contradicting itself, and the register is simply the first place anybody has seen both halves at once.

A note on scope: the pass only compares an organisation's shared memory. A member's own memory is not somewhere to raise a question for colleagues.

What it does not do

• It does not decide. A pair is two remembered statements, not a finding. Before anyone acts on one, the system that would settle it — the CRM for what a customer is, the ledger for a figure, the site for a public claim — is the thing to look at, and the tool says so to the model as well as to you.
• It only reads shared memory. A member's own memory is not somewhere to raise a question for colleagues.
• It will not raise everything. The nightly pass works out from the newest memories, within a band of similarity, ranked so that a model's attention goes to the pairs most likely to matter. A contradiction nobody has written down twice is not findable at all.
• An empty register is not a clean bill of health. Memtro reports it as a fact about the nightly pass, and offers to compare two systems instead.
• It cannot fix the business. Most of these pairs are not memory problems: a website that no longer describes the company is a contradiction in the company. The register is the first place both halves have been seen at once.

Also worth reading

• Memory — how facts arrive, and how history is kept.
• Jargon — the words your business uses in its own way, which is worth settling before contradictions are judged.
• Commitments — the same shape, for promises.