One typed statement, not a paragraph.
A substrate is not a document. It is a set of entries, and an entry is the smallest thing the graph is made of: one statement, of one kind, written by one author, linked to the entries it relates to. The kind is recorded as a field, so the same record that holds a decision announces that it is a decision without a reader having to infer it from the wording.
Typed, not prose. Every entry carries a type and structured metadata, so the model reasons on structure instead of re-parsing haze. A decision isn’t a paragraph that happens to contain one. It’s an entry of type decision, and the model acts on the type without re-reading the prose.
What an entry carries.
The Standard sets out a required core and two structural fields, links and metadata, that carry most of the value. The table below follows that definition.
| Field | Role |
|---|---|
| id | A stable identifier, never reused even after the entry is superseded. It is the reference other entries point at, so a link keeps resolving to the same statement for the life of the graph. |
| type | What kind of statement this is. A reader acts on the type without re-reading the prose: a decision is an entry of type decision, not a paragraph that happens to contain one. |
| content | The statement itself, written once, by the author who reached it. |
| author | Who wrote the entry, human or model. The field carries provenance, which is what a later reader weighs the entry by rather than taking its claims at face value. |
| created_at | When the entry entered the graph, fixing its place in the order of what was known when. |
| links | Typed edges to related entries: supersedes, replies-to, references, and others. An entry with no links is a leaf; the value compounds with the edges. |
| metadata | An open key-value space for the structured fields a given type expects, so a review can carry its verdict and a directive its status without either being buried in prose. |
Typed beats prose — the field agrees.
That structured memory outperforms free text is, by 2026, close to consensus in practitioner guidance. The reasons given are consistent across sources: a typed payload is compact, so it costs fewer tokens to carry; it does not have to be parsed for format on every read, so it introduces less room for error; and it is easier to retrieve accurately, because a field can be matched exactly where a sentence can only be matched approximately. The write-manage-read taxonomy in the 2026 survey of language-model memory (arXiv 2603.07670) names this the representational-substrate axis: what shape the memory takes, independent of where it is kept.
This is the efficiency case, and it is real. It is also not the whole argument, and treating it as the whole argument understates what typing does.
Typed links make relationships first-class.
The differentiated claim is not that typing is cheaper. It is that a typed link is a queryable structural claim, and some operations exist only once it is one.
Consider the request: show every decision this one replaced. Over free text, it has no reliable answer. A reader can search for phrases like “this supersedes” and hope the wording was consistent, but the relationship lives in prose, and prose is not something a query can trust. Encode the same relationship as a typed supersedes edge and the request becomes exact: follow the edges of that type, and the replaced entries are returned, with nothing to interpret.
The operation the whole specification turns on, supersession, cannot be defined without it. Neither can a causal chain: a claim that one decision exists because an earlier one turned out wrong is an edge of a particular type, not a hopeful reading of two paragraphs. The typed entry is the foundation both of those pages stand on.
One entry, field by field.
A single decision entry, read as its structure rather than its sentence.
type: decision · author: a colonel session · content: “Move from SQLite to Postgres once concurrent writers appear.” · links: supersedes the earlier SQLite decision, references the blocker that prompted it · metadata: the branch it belongs to and the directive it answered.
Nothing in that entry needs re-reading to be used. A later session selecting every decision on the branch finds it by type; a session asking what replaced the SQLite choice finds it by following one typed edge; a session weighing how much to trust it reads the author. The sentence is there for a person. The structure is there for the next reader, human or model.
What the rest of the substrate is built from.
The typed entry is deliberately small. What it makes possible is not: because the unit carries its own type and its own edges, the graph above it can support operations that a pile of text cannot. Memory the model reasons on rather than retrieves from depends on the relationships being encoded rather than described. The formal, field-level definition, including the exact required set and the link relations, is set out in the Vinculum Standard.
Common questions.
Why not just write a good paragraph?
A paragraph has to be re-read and re-classified every time it is encountered. A typed entry states its kind once, in a field, so a reader can select every decision, or every open question, without parsing text to decide what each record is.
Is the list of types fixed?
No. The reference implementation carries roughly ten first-class types in production use, and the set has grown as new kinds of coordination were needed. The point is that a type exists and is acted on, not that the vocabulary is frozen.
Isn't this just a database schema?
A schema constrains storage. The distinction here is what the typed edges make possible for a reader: operations defined over the relationships between entries, such as supersession, that free text cannot support reliably.