Chapter 19

How Memory Becomes Material

Chapter 18 resolved one Node. Finding one Node does not mean an Agent can answer.

What an Agent needs is the set of memories tied to the question, and those memories carry different kinds of evidence and relations. That selected, connected set is answer material.

1. brief is not another name for search

vinc_brief is not a search that returns a little more. It takes one question, runs the search of Chapter 16 and the resolution of Chapter 18, and assembles in one call the material needed to understand the resolved Node.

Search asks what is related. brief asks something else: which memories have to be seen together to understand this question. It is the deterministic assembly the desktop app already performed, ported and folded from a chain of tool calls into one.

2. The material centers on the resolved Node

Material gathers not around the question but around the resolved Node.

Material gathers around the resolved Node: its description, excerpts, relations with their evidence, and the grounds every claim stands on resolved Node description quoted, with where it came from excerpts which document, which section relations kind, direction, evidence, confidence, validity grounds every id the material stands on

Gathered this way, provenance can be followed: why a sentence was included, which Record or relation it connects to, which document it was drawn from.

3. Nothing is generated

There is no model inside brief. Every sentence it returns already existed somewhere.

The Node's description is a quotation, and the response says whether it came from the Concept's description column, from a property, or from a document excerpt. Excerpts carry their document id and section title. Relations carry the evidence sentence and confidence of Chapter 23 as stored. The grounds stay complete even when the result is shrunk to fit its budget.

An empty description column is not written into. The rule set in September 2026 is this: only a sentence already on the Node is promoted into the column, a line kept in its properties or an excerpt linked by its title. An excerpt linked only by provenance is not, because calling a sentence that shares almost no word with the title a description would contaminate the Node itself.

4. There are three statuses

The first line of a brief is its status, and a reader looks there first.

resolved. One Node won. The material is filled.

choose. The leading candidates tied. A candidate list comes back and no material. Asking again by the exact title usually settles it.

none. Nothing matched. If something matched and could not be read, that is reported separately, because the next move is different.

As Chapter 18 showed, the product returns choose on most questions that have an answer. A proposal to assemble material around the first candidate anyway exists, is built, and ships off. When the first candidate is wrong, that material would hand over the context of the wrong Node. In September 2026 the ruling was to leave it off.

5. More material is not better

Pull in everything connected and an Agent's judgment gets harder, not easier. The result enters the Agent's context, and that context is finite. So the result has a budget.

Over budget, it is reduced in a fixed order: excerpts first, then relations. It is not truncated but shrunk by policy, the grounds are left alone, and the response says it was shrunk, because a quietly shortened result is indistinguishable from a whole one.

Measured once, brief returned 5.5 times fewer bytes than a search chain for the same questions, and 3.2 times more of what it returned was actually used. Size and shape reach straight into call cost and the quality of judgment.

6. When material is thin, say so

brief's purpose is not to produce something every time. There may be no Node for the question, thin relations, or too few supporting documents. Then that state is the result; none and choose are how it is said. Hiding it behind a plausible answer is the opposite of what a memory system is for.

7. Answer material is a middle layer

Answer material is neither the final answer nor the unprocessed source. Source memory becomes answer material through search and resolution, and the Agent interprets it into an answer.

With this layer an Agent need not read every document and walk the graph from scratch each time. Without it, the Agent has to know the inside of the memory system and call several tools over and over.

One thing came out of measurement. Whether brief helps is decided not by the shape of the question but by the kind of the resolved Node. Kind is invisible from the question, so an Agent cannot route around it in advance. The numbers are in Part 4.

How this material crosses to an Agent outside the system is the next chapter.

Continue with Chapter 20, How Memory Leaves the System.