Statements: your decisions, in your words

Three months ago you wrote a meeting note: "We'll price Pro at $29/month." Five weeks later, another note: "Pricing update: Pro moves to $39/month." Today a teammate asks what Pro costs. Both answers sit in your notes, and search will happily hand you the stale one. What you want is a memory that keeps both sentences, knows they're both decisions, and knows the later one revised the earlier. It shows the old price struck through, with a pointer to what replaced it.

That's the statement layer. Entities and relations record the world your notes describe; statements record you. The vocabulary you just taught the extractor folds many verbs into one edge, and this layer does the opposite: it keeps whole sentences intact.

A Settings toggle called "Capture your decisions and questions" is on by default. With it on, extraction lifts your key sentences into the graph word for word. It sorts them into seven kinds, tracks speaker and subject, and links the ones that argue. It also keeps your revision history honest, which is where this lesson lands: the "Revisions" panel you'll use daily.

Why your exact words matter

When extraction reads a saved note, it also finds sentences where you commit to something: a decision, a question you're carrying, a claim, a quote worth keeping, a preference, a value, a stance. Each becomes a first-class node in the graph, a statement, anchored to the note it came from, addressed as that note plus a statement number. The stored text is your exact sentence, never a paraphrase. So when the app later answers "what did we decide about pricing?", it quotes you word for word.

• The exact excerpt: your sentence, unedited.
• A kind: exactly one of the seven below.
• What it concerns: links to the entities involved, but only the ones the same note asserted.
• Who said it: recorded only when the speaker isn't you.
• A polarity: favor, against, or neutral, on stance and preference statements only.

Seven kinds, and who said it

The seven kinds split into two families. Four are point-in-time acts, things that happened once, in that note, on that day: you decided, you asked, you claimed, you quoted. Three are standing dispositions, things that stay true about you until you revise them: what you prefer, what you value, where you stand.

Open any entity's profile page and the statements about it appear under "In your words", grouped with an icon each: "Decisions", "Open questions", "Claims", "Quotes", "Preferences", "Values", and "Stances". Each statement links back to its source note, and the section appears as soon as your notes carry statements.

Two of the standing kinds carry an extra bit. Stance and preference statements record a polarity, favor, against, or neutral, shown as a small badge next to the sentence. The polarity comes from a fixed set, so a badge always means one clear thing.

Every statement also knows its speaker. When the sentence is yours, nothing extra gets recorded: absence means the note's author. When your note quotes or reports someone else, like a quote's source or "Sarah decided X", the statement records that speaker and points at that person's entity. The entity page then shows the speaker's name after the sentence.

When statements argue: supports, rebuts, respondsTo

Notes don't only state facts. They argue: "Because of that…", "this contradicts…", "answering yesterday's question…". When your text explicitly links two statements, the graph records the link as supports, rebuts, or responds to. One rule keeps this trustworthy: these links exist only when you linked them in the text yourself. The machine never infers your reasoning, so a chain of supports and rebuts is always one you wrote.

In the note:

  1  Our churn data says the first week decides retention.      → claim     (statement 1)
  2  Because of that, we'll rebuild onboarding before billing.  → decision  (statement 2)
  3  Minsu disagrees. He says pricing, not onboarding,          → claim     (statement 3)
     drives churn.                                                 (said by Minsu)

Stored links, only because the text itself connects them:

  statement 1  supports  statement 2     ("Because of that")
  statement 3  rebuts    statement 1     ("disagrees")

No explicit link in the text → no link.
The graph never invents your reasoning.

Changing your mind: supersedes, retracts, derived status

Minds change, and two more statement-to-statement relations record it, again only when your text says so. Supersedes: this statement replaces an earlier one. Retracts: the earlier statement no longer stands, and nothing takes its place. From these links, every statement gets a lifecycle status: current, superseded, or retracted. That status is never written into the store. It's worked out at read time from whatever revision links exist right now.

You'll see the result all over the app. A superseded or retracted statement renders struck through with a "Superseded" or "Retracted" badge. Next to it sits a link, "Superseded by {note}" or "Retracted by {note}", that opens the revising note. The old words are never deleted, just kept and honestly labeled.

When a later note revises an earlier decision

So far, revision meant linking statements yourself, inside one note. But real decisions usually change across notes: June's meeting note decides one thing, July's decides otherwise, and neither mentions the other. After extracting a decision-bearing note, the app can infer that a later note revises an earlier note's decision about the same entity. These machine guesses never touch your notes' graphs. Each proposal sits in a separate `#analysis` named graph alongside them, stamped with a review status and the time it was proposed, waiting for your review.

• Only decision-kind statements take part. This is the layer for how your decisions evolve.
• Ordering uses the date the note itself states, in frontmatter or prose; the file's modified time is never used. Date your decision notes and they line up on their own.
• High-confidence proposals auto-apply and arrive marked "Applied", and they always stay reversible with "Undo".
• A rejection is remembered forever, so the same pair is never proposed again.
• Only confirmed links make a decision render as "Superseded" across the app.

Statement

Your verbatim sentence stored as a first-class node, addressed as the note plus a statement number.

Statement kind

One of the closed seven: decision, question, claim, quote, preference, value, stance.

Stance polarity

favor, against, or neutral. A closed list, only on stance and preference statements.

Attribution

The recorded speaker when it isn't you; absence means the note's author.

Rationale link

supports, rebuts, or responds to. Statement-to-statement links, author-explicit only.

Supersession

A later decision revising an earlier one, within a note or inferred across notes.

Derived status

current, superseded, or retracted. Computed at read time from revision links, never stored.

Note date

The date a note states about itself, in frontmatter or prose. Revisions order by this date; the file's saved time is never used.

Wrap-up: a memory that argues honestly

Statements complete the three-layer model: entities give your notes identity, relations give them structure, and statements give them a voice and a timeline. So when you ask the agent "what did we decide about pricing, and when did it change?", this layer answers in your own sentences, attributed and honestly revised. Next you'll learn the commands that call on it directly: /revisions, /tensions, and /brief all read the decisions and stances you captured here.