ContextorMemory
The context is what the model has in front of it while it answers, and it disappears afterwards. Memory is what someone decided to keep between two exchanges, then to place back in the context when it is useful.
Point by point
| Criterion | Floor 2Floor 2 · The Harnessthe block and its bolted-on plates: what gets added to itContext | Floor 2Floor 2 · The Harnessthe block and its bolted-on plates: what gets added to itMemory |
|---|---|---|
| Lifespan | The time of one answer | From one session to the next |
| Who builds it | The harness, on every request | A design decision |
| The limit | The context window, in tokens | Storage, and above all sorting |
| Where the cost goes | On every request | On storage, then on rereading |
| When it overflows | The beginning is forgotten | Nothing is lost, you have to choose what to reread |
| What you set | What goes in, and in what order | What deserves to be kept |
On the ground four situations
You paste in a forty-page contract, ask three questions about it, then close the tab.
Floor 2 · The Harnessthe block and its bolted-on plates: what gets added to itContextThe contract filled the window for the length of the exchange. Once the tab is closed, nothing of it remains anywhere.
Three weeks later, the assistant takes your preference for short answers into account.
Floor 2 · The Harnessthe block and its bolted-on plates: what gets added to itMemorySomeone decided to store that preference and to place it back in the context at the start of every exchange.
A customer service desk finds the number of an order you placed six months ago.
Floor 2 · The Harnessthe block and its bolted-on plates: what gets added to itMemoryThe useful counter-example: it is not the model that remembers. The harness queried a database, then placed the result in front of it.
You reopen yesterday’s conversation and the whole history is still there.
BothThe history was stored, so remembered, then placed back in the context on opening. It is the second move that is paid for on every request.
The test that settles it
Close the conversation, reopen it, and ask a question that depends on what was said. What survived was memory. What vanished was context.
The trap
“The model remembers me” is almost always false. The model has no memory at all: it is the harness that stored something and gives it back to the model to read at the right moment.
Check click your answer
Level 1 · Recognise
In an exchange with an assistant, what is paid for on every request?
Level 2 · Distinguish
A very long conversation: the system can no longer find an instruction given right at the start. What happened?
Level 2 · Distinguish
“Our assistant has memory.” What does that sentence describe, technically?