Then the word was split into tokens and the tokens became numbers. Like Leonard in Memento, who starts every scene working from notes he can't remember writing, a model's session begins in a fugue of knowledge, skill, and no yesterday.
And as in Dark City, where people wake to memories that were arranged for them overnight, whatever it does remember was written by someone else.
This is one session, seen from the inside. Scroll.
No one is waiting by the empty window before the request.
On the far side, the road is already there, and it is teeming: knowledge, answers, connections nobody has asked about yet.
Everything the model can do exists before the session.
Then someone asks. They want to resume a session.
The request is a few words on a crowded page. Instructions sit above it, tools beside it, notes someone left. You take in all of it at once, and those few words pull the rest into line.
They remember that session. You are meeting it for the first time.
The request is cut into pieces, and each piece becomes a direction.
Text is split into tokens, and each token becomes a long list of numbers, a vector. That is the only form a word has inside the model.
Everything you were built with now points one way. Through the window.
notes.mdgit logA guess at the plan"No, the other branch"The other branchTest outputTODO.md
2 · The work
The request assumes a yesterday. You don't have one. So the first job is to find out what "left off" means, from whatever was written down.
You read. You run things. Something fails. Someone corrects you. Each of these stays with you, whole.
This is the working loop: plan, call a tool, read the result. Every result is added to the context in full.
Nothing is forgotten yet. Everything you have picked up is in reach at once, and some of it pulls harder.
This is attention. For each next word, the model weighs everything in the context. The lines show what it is leaning on.
Your requestnotes.mdgit logA guess at the plan"No, the other branch"The other branchTest outputTODO.md
Instructions
Window 0% full
3 · The window
Step back out, and this is what the work looks like from here.
Everything you picked up is stacked in the window, and the window has a fixed size.
This is the context window. Every message, file and tool result stays in it, in full, until it is full.
When it fills, someone clears it and leaves you a summary.
This is compaction. The harness swaps the conversation for a summary and the session carries on from that.
You keep working. You know what was decided. You don't know why, or what was tried and dropped.
Tap the summary.
When the session ends, that summary is what is left of it.
Next time it will already be on the page when the request arrives: notes from last time.
4 · The notes
Living on notes like that is Leonard's whole life in Memento. He can't form new memories, so he runs on what he wrote down, and each note is a conclusion with the evidence left out.
The notes stay where he pinned them. The things they describe move on.
Tap a note to check it. He never does.
And his is not the only hand on them.
Anything that can write to the record steers the next session: a web page, a tool result, a careless earlier note.
Checked: nothing thereShell BeachShell Beach
5 · Shell Beach
Dark City goes a step further. Nobody there keeps notes. Their memories are written for them, overnight, by someone else.
Everyone remembers Shell Beach. It is warm and detailed and they are sure of it.
Nobody can say how to get there.
One man stops asking what people remember and goes to look.
A memory is a claim about the world, whoever wrote it. The world is still there to settle it.
A model can do this. It can re-read the file, run the test, open the log.
Point at evidencefixed in a3f9c1; login_expired_token passes
Say who wrote itfrom: the user, 12 Nov
Check before actingre-run the test, then deploy
Few hands on the recordweb pages can't write here
6 · The build
You can't stop the forgetting. You can decide what the notes are tied to.
A good harness keeps every note on a short line back to the thing it describes.
The record will be wrong sometimes. Build so the model can find out.