A novel is bigger than any model’s memory

A language model reads a fixed amount of text at once. A short story might fit. A novel does not — not the manuscript, not your notes on it, not the eleven things you decided about Tom in chapter one. So every writing assistant has to choose what the model sees, and most of them choose by accident: whatever is nearest the cursor, and whatever you remembered to paste in.

ProseSpun chooses on purpose. Three things do the work.

Cards are the memory. Every person, place and thing your story names gets a card: a summary, a few traits, the other names the prose uses for it. When you ask for a sentence about Tom, Tom’s card comes with the request — that he is never late, that he works nights. You don’t keep the cards up to date by hand. ProseSpun reads what you write and proposes changes, and you say yes or no.

Summaries roll up. Each scene gets a short summary once you’ve moved on from it. Scenes roll up into chapters, chapters into the story. When you write in chapter twelve, chapter two arrives as a paragraph, not as twenty pages. The model knows what happened without being asked to read all of it again.

Facts have a place in the story. A character’s broken arm, a promotion, a house that burns down — these are true from a particular scene onward. ProseSpun keeps each change pinned to where it happens, so when you go back to revise chapter three, the suggestions know the house is still standing.

The point is not to remember everything. It’s to remember the right things, as of the page you’re on.

One part of ProseSpun puts all of this together for every request: your Setting, the cards in the scene, the summaries behind it, the paragraphs around your cursor. Each part has a limit, so a long cast can’t crowd out the scene you’re writing.

None of it is visible while you write, and none of it changes your page. Whatever comes back is a proposal, with the changes marked, waiting for you to accept it, discard it, or try again.