Chapter 4

What Should We Remember?

Chapter 3 showed that memory changes the AI's answer. Our definitions, decisions and experience become part of the conditions that shape it.

So what should we remember?

Store everything? No. Storing everything is not the same as making memory. Pile it all up at once and it gets harder, not easier, to find what matters.

People do not remember every moment with the same weight either. Some events stay for years and some remarks vanish. Some decisions are remembered because they change so much of what follows, and some experiences become the yardstick for understanding later ones.

A digital system is no different. Start with what we may need again, what helps us understand other memories, what can be checked against evidence, and what can still be followed as time passes.

Good memory is not the memory that stores the most. It is memory stored in a form that can be understood again.

1. Before remembering, we have to tell things apart

Suppose we want to keep a memory about a place. There is the name of the place, an event that happened there, a person who took part, a decision made at the time, and perhaps a document explaining the decision.

We could store all of it as one long text. But then asking questions later becomes hard.

“Which events are connected to this place?” “What was decided in that event?” “Where is the evidence for that decision?”

To ask, we first have to tell apart the things we are remembering. That distinction is what gives memory its structure.

2. Naming something is cutting up the world

Calling something a concept is not choosing a name for it. It is deciding to distinguish it from everything else.

With the concept “Decision” we can tell an ordinary sentence from a decision. “We chose B” may be part of a record. But if that sentence marks a real choice, we can remember it as a Decision.

Then the question changes. Instead of finding a sentence, we can find the decision. From there we can follow the reasons, the evidence and the related records.

This is why concepts are needed. A concept is not a folder for filing memory. It is an agreement about what counts as one meaningful thing.

3. Where structure begins

As these distinctions accumulate, a structure appears.

There are people, events, decisions, documents and concepts, and they stand in relationships to one another. A person makes a decision, an event triggers one, a document explains one, one decision leads to another.

At that point we hold more than a list of concepts. We hold a model of what exists and what kinds of relationships those things can have. In Vinc that model goes by the plainest names it has: the kinds and the relations.

Building models of what exists is an old undertaking and a far larger one than any single product. Vinc sits inside that tradition rather than standing for it. What it holds is one small working model for one graph: the basic agreement of distinctions and relationships used to give memory structure. Not everything stored in Vinc, and not a dictionary trying to classify the whole of reality.

4. Concepts and records are not the same

One more distinction matters here. A Concept and a Record are not the same thing.

A Concept is a lasting reference for what we mean. A Record is evidence actually left at a particular moment.

Suppose a team uses the concept “canonical record”. There is an explanation of the concept, and there are the documents and decisions where it was applied in a real project. The concept holds the meaning steady; the records keep the trace of that meaning in use.

Keeping the two apart makes something important possible. We can remember both the meaning we use now and the original record that supports it. When an AI answers about a concept, it can pass on more than the definition; when needed, it can follow the definition back to where it came from.

5. Decisions have to be remembered too

Not every record is the same kind of memory. The kind that keeps mattering long afterward is the Decision.

Remembering only that we chose B is different from remembering why. If the decision changed later, the change has to be remembered too, or an old decision gets mistaken for a current rule.

That is why time matters for decisions. When it was made, on what grounds, whether it still holds, and what changed it must stay connected before a past judgment can be understood in the present.

Memory preserves not only the old answer but when, and why, that answer held.

6. When structure appears, the graph comes alive

Chapter 3 said a graph preserves the relationships in memory. Now the reason is clearer. To build a graph we first decide what counts as a node and what counts as a relationship. That is the job the kinds and the relations do.

A meeting (Episode) is about a Decision. The Decision is about the Concept it settles and derives from the decision before it. A Record, such as a comparison run, attaches to the Decision with its evidence. Vinc has six kinds, these four plus Bridge and Entity, and six relations: relates_to, prerequisite_of, derives_from, inherits_from, about and connects_via. Chapter 11 covers what each one means.

Here the graph is not a technique for drawing lines. What we decided was worth remembering, and which relationships we judged they have, is built into its structure.

The graph is the shape of memory. The kinds and the relations are closer to the vocabulary and the agreement that give it that shape.

This is what lets Chapter 3's conditioning work. What reaches the AI is not a random pile of documents but a structured part of memory we judged worth keeping, and it can be followed back to the original records.

7. A good model does not try to explain everything

One temptation has to be resisted here: classifying all of reality. Aim at a perfect model and the model itself becomes the project.

But the purpose of a memory system is not to reproduce reality. It is to keep the meanings we will need again from being lost.

So we distinguish only as much as we need. Add a new kind of memory when it is called for, stop using a distinction when it no longer means anything. What matters is not how elaborate the classification is, but whether the structure makes memory more accurate and lets us explore it again later.

8. When many hands begin to write

We now have a structure. We decide what to remember, define it as concepts, connect records and decisions, and express the relationships as a graph.

For a memory used by one person, that might be enough. But Vinc's graph is not filled by one person alone. People write records, an AI assistant proposes, an Agent finds new connections, and someone else brings in their own memory.

What happens when the same thing starts being written differently? When the same distinction goes by different names, and the same relationship gets drawn in opposite directions?

The moment memory passes through many hands, distinctions and relationships are not enough. We need shared rules for writing them down.

That is the next chapter.

Continue with Chapter 5, Memory Needs a Grammar.