Chapter 5
Memory Needs a Grammar
Chapter 4 ended with a question: what happens when many hands begin writing into the same graph?
A memory used by one person can be written however that person likes. But Vinc's graph is a place where people write records, an AI assistant proposes, an Agent discovers new connections, and someone else brings in memory of their own. Once each of them writes a little differently, the graph quickly becomes a place where several languages are mixed together.
That is why memory needs a grammar.
1. Write the same thing differently and memory splits
Suppose two people record the same event. One calls it an Event, another an Episode, a third simply leaves a Record.
Relationships go the same way. One person writes that a decision happened because of an event; another draws the relationship in the opposite direction.
Taken one at a time, every record looks reasonable. The moment they connect, the trouble starts. The person reading the graph, and the AI reading it, interpret the same meaning differently.
Once memory is shared, personal ways of writing are not enough. There has to be an agreement everyone keeps.
2. The kinds say what; Convention says how
Chapter 4's kinds and relations divide up the world we remember. They say what a Concept is, what a Record is, what a Decision is, and which kinds of relationship can exist.
But the kinds alone are not enough. How is the same Concept actually written down? How is the direction of a relationship expressed? How large should a new node be? What happens when a duplicate memory turns up?
These questions call for Convention: the shared rules of writing.
The kinds and the relations are the vocabulary of memory. Convention is its grammar.
Keeping the two apart also explains why the graph is more than a database.
3. Machines have to read the grammar too
If only people worked together, the rules could be written up in a document and agreed. But in Vinc, AI and Agents read and write the graph directly. So the rules have to be readable by a machine as well.
“What does this graph call a Concept?” “Which edges are allowed?” “Which areas may be generated automatically?” “What must a person confirm?”
If a person had to answer these every time, the same knowledge would wobble each time the Agent changed.
So Vinc keeps a document that describes the graph's own rules and state: GRAPH.md. People can read it as a guide, but it is closer to a contract, written so a machine can read it, stating the rules under which the graph exists.
GRAPH.md does not stand in for the memory itself. The canonical record is the graph and the records connected to it. GRAPH.md says how that memory is to be read, how it is to be added to, and by which rules it is to be connected.
That is why GRAPH.md has two zones. Rules that people set and people are responsible for, and information such as the current state of the graph that the machine computes and refreshes on its own. The reason for keeping them apart is simple. A judgment a person must answer for and a state the system can reproduce are different things.
4. Not every rule has the same force
When the rules of a graph are written, they should not all be bound with the same strength.
Some rules the system sets, such as the schema and basic integrity: break them and there is no graph at all. Some rules the owner of a particular graph sets. Some change with the subject or the situation.
Schema-fixed. Rules the system and its structure must keep.
This graph decides. Rules a particular graph sets for its own purpose.
Varies by topic. Rules that change with subject or situation.
The distinction matters so that an Agent does not mistake every sentence for a command of equal weight. The structure of memory holds both what must be kept and what may be chosen.
5. What can be checked is different from what must be judged
Another principle follows. A checker does not judge meaning.
Is the format valid? Are the required sections present? Are the boundaries of the generated zone intact? Do the references exist? These the system can check.
“Is this really a good Concept?” “Are these two records really the same memory?” “Does this decision still hold?” These are different. They are judgments about meaning and context.
So Vinc keeps verification and judgment apart wherever it can. The system reports broken structure. The AI proposes merges or new relationships. But a change of meaning is confirmed and approved by a person. The principle from Chapter 3, show the conflict and never merge on a guess, becomes a rule here.
An Agent may propose a memory. It must not decide, on its own, that the memory is true.
6. Grammar makes memory repeatable
Chapter 3 said that what we show the AI conditions its answer, and Chapter 4 said that the material has to be structured memory. Now one step further.
For structured memory to keep the same meaning over time, it needs a grammar for writing it.
If a Decision written today and a Decision someone else writes months from now follow different rules, the two cannot be used reliably in the same graph. Keep a shared convention and memories stack. One record sits on another, one decision leads to the next, and different people's experience connects inside the same structure.
That is repeatability. The same memory can be read again by the same rules, and new memory can be added within the same structure.
7. The grammar of memory is, in the end, for people
A rule document like GRAPH.md looks as though it exists for machines. What it finally protects is human memory.
When the person changes, when the assistant changes, when the Agent changes, when time passes, the meaning of memory must not drift by accident. So in Vinc, rules are not a coding style. They are the social and technical agreement that lets memory continue in the same structure no matter whose hands it lands in.
Once memory can be distinguished, connected, and kept under a shared grammar, the next question follows on its own.
Where is that memory?
The next chapter looks at memory as a space. It has position, distance, and time.
Continue with Chapter 6, Memory Needs Space.