How to write a novel with AI: what breaks

Why does ChatGPT forget my novel halfway through?

A chat assistant does not remember your book. It re-reads the conversation from scratch before every reply, and what it can read in one go — the "context window" — is large but finite. Once the thread outgrows it, the earliest messages stop being sent: nothing was deleted, but by chapter fourteen, chapter one is no longer in front of it.

Your cast, your settings and your outline are messages, not records — so when they stop arriving, the model fills the gap with something plausible: a sister becomes a brother, a scar moves to the other arm. Not a malfunction; a system answering without the page it needed.

Would a better prompt fix it?

Partly, and only for a while. A well-built prompt — a clear brief, the cast list, what the scene has to do — genuinely improves the next few pages. It does not change what happens as the thread gets longer: that is a limit on how much text can be read at once, not on how well you asked.

And it is not a quality problem. The models inside general chat assistants are the same class a dedicated writing tool uses, often literally the same ones. The gap is not intelligence: a chat is a general instrument, and a novel is a long structured document with hundreds of moving parts that has to survive months of work.

So the fix is not a cleverer message. It is somewhere to keep the story that is not the conversation — where your cast, your settings, your structure and every scene exist as things in their own right, and the right ones get handed over on every reply.

What does Bloomwright do differently?

The story is stored, not remembered. Your characters, places and the shape of the book are entries in a project, not messages in a thread, and before each reply the assistant is handed the current state of that project automatically. You never re-paste your outline, and a detail from chapter two is still true in chapter thirty.

You work scene by scene, and every scene keeps its own history. Each scene has its own page and a short brief — what it has to do to the reader — and you can draft it a piece at a time, or rewrite the one paragraph that is not landing without touching the three around it. A restore point is saved before every rewrite, so the version you preferred is still there.

The structure is something you can look at, not a paragraph you keep re-describing: the sequence of turning points your book has to hit, in plain words rather than craft jargon. You can read the whole thing back as a finished book and export it as EPUB or Markdown. Bloomwright proposes and you decide — the book is yours, and we never claim to have written it.

What you write in Bloomwright stays yours, and the terms of service say it in those words. We claim no ownership of it and no licence to publish, sell or reuse it, and nothing you write is shown to other users. We do not use your writing to train AI models, and neither do Google and Anthropic, whose models read your project to answer you: both state that what we send them is not used to train theirs. Prose is generated only through providers that, once the request is done, keep neither your text nor what they wrote.

It is free during the open beta, which runs until 30 November, with a daily limit on the AI; what happens afterwards hasn't been decided. You don't need an account to start — the first conversation about your story runs in your browser. All you need is your imagination.