Jargon

Every company runs on a private vocabulary. "Settlement", "MID", "exposure", "active merchant" — words that mean something particular here, and that nobody has written down, because to the people using them the meaning is obvious.

It costs twice. A new joiner spends weeks inferring it from context. And an assistant answering from your memory reads those words in their ordinary sense, which is how a confident answer comes out subtly wrong.

Dashboard → Memory → Jargon finds them and asks what they point at. Jargon is off until you turn it on, under Organisation → Features.

How a word gets found

The nightly pass reads your shared memory — the text, not the vectors, so it is the cheapest of the nightly passes — and shortlists words that recur in at least three memories across at least two systems. That second test is the whole trick: a term used only in one feed is that feed's quirk, while a word that turns up in the CRM, a meeting transcript and the mailbox is shared vocabulary.

Ordinary English is skipped, as are the acronyms everybody knows (API, VAT, GDPR) and anything that is the name of something in your graph — a company or a person is an entity, not vocabulary.

A model then says which of the shortlist is genuinely your own word: an acronym particular to your business or industry, an ordinary word used as a term of art, or the name of an internal process, document, tier or stage. It writes what the word points at from the uses alone — no guessing from the word's shape, no general knowledge — and where the uses do not say, it leaves the note empty and says its confidence is low, because a term worth asking a person about is a perfectly good answer.

The most valuable thing it can report is disagreement. Where the uses plainly carry more than one meaning, each is listed as a sense, and the row is flagged: a word two teams use differently is the one that costs a misunderstanding.

A pointer, not a definition

This matters enough to be the feature's main rule.

A note is written from a handful of uses out of thousands, and these words almost always reach further than a sentence about them. So nothing here is treated as a closed definition. The model is asked for the broader reading where there is a choice, and to describe what a word is getting at rather than draw a boundary around it. The page says the same to you: broad beats exact, and the button adopts it as a pointer.

What Memtro does with one follows from that. A note stops a word being read in its everyday sense — and it is never used to narrow what somebody means, to tell them they are using their own word wrongly, or to rule something out because the note did not mention it. Where an answer turns on the edge of a meaning, Memtro is instructed to go and look at how the word is actually used and say what it found, rather than lean on the note.

A note that is too narrow is worse than one that is vague, because it gets read as the whole of the meaning. Which is why widening an entry is as useful as correcting it: finding that a word reaches further than its note is exactly the thing worth recording.

Adopting one

Each row shows the word, how widely it is used, one example from each system it appears in, and the senses where the uses disagree. The note arrives pre-filled with what the model read out of those uses, and is editable — the value of the register is your words rather than a model's guess, and the fastest route to your words is a draft you can correct.

Adopting it does two things:

1. It becomes a memory. "In this organisation, X refers to Y — a note on how the word is used here, not a closed definition; it may cover more than this." So it is retrieved like any other fact, turns up in memtro_search, and carries its own caveat wherever it is read.
2. It joins the prompt. The confirmed glossary — the widest-used twenty-five terms — rides in the system prompt of every request, so answers start from your usage instead of the dictionary's. A model's own proposal never joins it: the prompt teaches what people confirmed, not what a model guessed.

Not jargon leaves a word out for good. Correct or widen changes a note in place. Ask again withdraws it and puts the word back in the queue — and withdrawing takes the statement out of memory too, because a meaning nobody stands behind should not be retrievable.

From chat, the API and MCP

• memtro_glossary looks a word up before answering a question that turns on it. Notes come back as what each word points at, with the pointer rule attached, and entries nobody has confirmed are marked as exactly that.
• memtro_define_term records what somebody just explained in a conversation — "a MID is the acquirer's merchant identifier" — raising the term if it was not in the register yet.
• memtro_update_term corrects a note, widens it, takes a word out of the glossary, or puts one back to be asked about again.

All three refuse to record anything the model merely inferred. A glossary of guesses is worse than none.

What it is for

Three things, in rough order of value.

Answers that use your words. Every reply, retrieval and nightly judgement after a term is adopted reads it your way. It makes everything else Memtro does slightly more accurate, including the contradictions pass, which reads the same words.

Onboarding. The glossary is the document nobody ever writes, assembled from what the company actually says rather than from what somebody remembered to document.

Noticing the ambiguity. The terms flagged as used more than one way are findings in themselves — two teams with two meanings for one word, which no amount of documentation would have surfaced because each team thought the meaning was obvious.

What it does not do

• It does not define anything. Every note is a pointer to what a word is getting at here, written from a handful of uses out of thousands. Nothing in the register is treated as a closed definition, and widening an entry is as useful as correcting one.
• It does not correct your people. A note is never used to narrow what somebody means or to tell them they are using their own word wrongly. Where an answer turns on the edge of a meaning, Memtro is instructed to go and look at how the word is actually used and say what it found.
• It will not find a word used in one place. The test is at least three memories across at least two systems, because a term used in one feed is that feed's quirk rather than shared vocabulary.
• It never records what a model guessed. Only confirmed entries join the prompt, and all three tools refuse to record anything inferred rather than said. A glossary of guesses is worse than none.
• It does not read your documents for definitions. It reads how the words are used, which is usually more honest than what a wiki page claimed three years ago.

Also worth reading

• Memory — how facts arrive, and how the prompt is built.
• Contradictions — things your memory believes twice, differently.
• Commitments — who promised what, and whether it landed.