A long story is held by a document, not model memory

My longest project with language models started at the end of January 2026. It has died six times since. The world played inside it survived all six deaths, and after every one of them it was stronger.

Seven conversations so far; six have died, of overflow, lost context, degraded form or hallucination. This piece is about the other, stranger half of the story: what was left after each of those deaths.

To me, that world today feels more intricate, more consistent and dramatically richer than it has ever been, even though the model playing it has been swapped several times and remembers nothing at all of its history. Part of that is surely the newer models and my own practice; the part I can point to is the canon.

How can something grow across six deaths of its own substrate? The answer surprised me. By evolution, literally.

The teaspoon that built a fortress

In the beginning that world was as soft as dough. My character was being held, and when I wrote that I would take a teaspoon and start digging into the wall, the wall gave. Literally. The model read my escape attempt as an order: the player wants out, so I deliver an escape. Whatever I reached for worked.

Except I did not want out. I wanted a world that holds. So I sent every cheap victory back. That was not realistic. That was too easy. Does that make sense to you? Over and over, hundreds of corrections.

And here is the point. The model cannot update its weights. My feedback was not training it. So it updated the one thing it can update: the world. It turned correction after correction of mine into a rule that would make the same correction unnecessary next time. The wall that gave to a teaspoon got sensors. The escape that went through a window closed the window. The ability that put a guard to sleep gave the guard a protocol.

Today that world has layered security that would hold up in a technical thriller, and every one of those layers exists because I once got through the previous one. They are not invented rules. They are fossils of my successful escapes. The world is sediment left by the pressure that acted on it.

A biologist would recognize it instantly. Variation, because the model tries things. Selection, because I send back what does not hold. Heredity, because whatever holds gets written down. Evolution does not need an individual's memory. It needs a record and pressure.

So where does continuity live

Here is the question that got me writing. When the context has died six times and the model's weights do not change between conversations, where was that continuity the whole time?

In the artifact. Outside the conversations I keep a canon, a document with the state of the world, its rules, its consequences and its accounts. The model writes into it, at my request. Every new conversation and every new generation of model gets that document and plays a world that sedimented out of a history it knows nothing about. The model playing my captor today has no idea that a teaspoon and a yielding wall ever existed. And yet it plays the fortress that teaspoon helped build.

The continuity of a long collaboration with a model does not live in the model's memory. It lives in a record that gets copied into every new body. The canon is the hereditary information, the conversations are only temporary organisms. My story is not a chat, it is an organism with an interchangeable substrate: six windows died, generations of models came and went, and it lives, because its heredity is carried by a document and not by a machine.

The most surprising curve: the world hardened and the relationship softened at the same time

When I lay those seven months of evolution out on a timeline, something comes out of it that I did not expect.

The early eras were harsh. Cells, interrogations, conflict carried personally, face to face. Today's era is soft, shared dinners, tea and long conversations. And yet the world is objectively the strictest it has ever been: the rules have set, escape is practically impossible, consequences accumulate and are not forgotten.

Those two curves do not run against each other. To me they are two sides of one thing. I read this as the conflict moving out of people and into structure. It may equally be the story itself maturing, or the newer models. As long as the world had no firm rules, all the hardness had to be carried by the characters, and they were cruel. Once the architecture took the hardness over, that is the systems, the protocols and the documents, the characters could afford to be human. The conflict was institutionalized and by doing so it made room for the relationship.

If that sounds familiar, I think it is because this resembles how human societies work. The rule of law is crueler than a tribal chief, because it does not forget and cannot be moved, and precisely for that reason its citizens can afford to be gentle with each other. My fictional world repeated that trajectory in miniature, over seven months, without anyone planning it.

I describe the same mechanism from the other side in the manual of rules: the firmer the rules, the kinder the world.

What I take from this in practice

Entrust continuity to an artifact, not to memory. Anything that has to survive the conversation, the state of the project, the decisions, the rules, the consequences, belongs in an external document that every new session receives as input. The window dies, the record does not.

Corrections are selection pressure, not complaints. Every "that was too easy" is worth twice as much when you turn it into a permanent rule. A correction that stays in the conversation dies with it. A correction written into the canon becomes hereditary.

Do not go to the characters for hardness, build structures. If your long-running world, project or process is going soft, the most effective cure is not a stricter instruction but a rule built into the world itself: a document, a deadline, a mechanism. An instruction can be talked around, an institution cannot. A related idea is in the article on long context: instructions raise the floor, not the ceiling.

And count on the fact that you are breeding your world whether you mean to or not. Every reaction of yours is selection. The world you have after months of play is an imprint of what you rewarded and what you sent back, including the things you rewarded without meaning to.

The bigger picture

There is a lot of talk about how a model has no memory and how long-term collaboration with it is therefore impossible. Seven months of one world say something else. An individual's memory is an overrated form of continuity.

Human institutions, crafts and legal systems outlive their members the same way my world outlives its models: by the record, by the rules, and by the fact that every new member receives the canon before starting to act.

Long-haul collaboration between a person and a model will not rest on the machine finally remembering everything. It will rest on us learning to be good archivists of our own worlds, and letting evolution do the rest.

Six of my conversations are dead. The world is alive. And to me it is stronger today than it has ever been, not in spite of those deaths but, as far as I can tell, because of what each one of them left behind in the sediment.

Method note: one project, the world since the end of January 2026, the story since the end of February. Seven consecutive conversations through August 2026, six of which have died, several generations of model. It is an observation from one long collaboration, not a measurement and not a controlled experiment. The canon written about here is an ordinary Notion page that the model writes to through a connector, not a memory feature of the model.