01 — part 1, conceptsconcept

What a memory is, and what it replaces

Memory engineering is the practice of deciding what an agent should write down, what it should re-derive, and what has to be true before a written record is trusted again.

The definition, and the alternative

A memory is a record written during one session and read during a later one. That is the whole of it. The word covers a great deal of machinery, and all of the machinery exists to serve that one sentence.

The important part of the definition is not what a memory is. It is what a memory replaces: going back to the source and working the answer out again.

	reading a record	re-deriving from the source
cost	one lookup	reading the documents, every time
correct when the world is still	yes	yes
correct after the world moves	only if something noticed	yes, always
can answer what did we say last quarter	yes	no

Every row but the last favors the source. That is not a rhetorical setup — it is the arithmetic this volume spends Parts 2 and 3 on, and the last row is what Part 4 is about.

Both routes answer the question, and the short one arrives first — that is the entire appeal of a store. Then one document is edited. Only the long route's answer changes. The record is not wrong because anybody made a mistake; it is wrong because it was written before the edit and nothing has looked since.

Nothing in this scene is measured. It is the trade being drawn; step 4 prices it.

So a store is a trade, not an improvement

The framing almost everything written about this subject uses is that memory makes an agent better. It is worth being precise instead: a store buys you the re-derivation cost and sells you the risk of being confidently out of date.

Whether that is a good trade depends on two things you can measure and one you have to decide:

#	what to establish	who supplies it
1	how expensive re-derivation actually is for you	measurement
2	how often your sources change	measurement
3	what a confidently wrong answer costs you	nothing here — you decide

Re-derivation is answering the question again from the original documents, as though the agent had never seen it before. In this volume it is always available and always correct, which is a property of the fixtures and not of the world — step 12 says so plainly, because every conclusion here rests on it.

What this volume does not treat as memory

Two things get called memory and are not this subject.

The context window is not a store. It is a working area that is discarded at the end of the session, and it belongs to the harness volume. Step 2 defines it and then leaves it alone.

A transcript is not a policy. Keeping every message an agent ever exchanged is a recording, not a memory system: nothing in it has been decided to be worth keeping, and nothing in it can go stale, because nothing in it claims to be true. The moment you extract a claim from a transcript you have built the thing this volume is about — and step 5 measures exactly that moment.

Where this volume starts. Not at "should my agent have memory" — at "what does answering this question again actually cost me". Every step after this one is an attempt to put a number on the other side of that trade.