In chapter three, her brother dies. In chapter seventeen, the AI has him pour the wine at dinner. If you've typed "AI forgets my story" into a search bar after a moment like that, you're far from the only one. It happens with ChatGPT, with Claude and with most other models. The prose can be perfectly fine. The problem is how the model sees your book: in pieces, and only the piece in front of it. Romance writers feel it hardest, because a slow burn is a promise the model has to remember: see the AI romance writer comparison.
Why AI forgets my story (and everyone else's)
It only sees what's in the window
A language model has no memory of your book between requests. What it has is a context window: the text sent to it right now. If a fact isn't in that window, it doesn't exist as far as the model is concerned.
Windows are big now. Anthropic's help center lists context windows from 200K up to 1M tokens on paid Claude plans, depending on the model. That sounds like enough for a whole novel, and in raw size it often is. The trouble is everything else in the chat: every draft, every rejected version, every "make it darker", every tangent about the magic system. The book is in there somewhere, mixed in with all the rest.
Long inputs aren't read evenly
Researchers studying long contexts found that models tend to use information at the start and end of a long input better than information in the middle, and that performance drops as the input grows. The paper is "Lost in the Middle" (Liu et al., 2023), and the effect showed up even in models built for long contexts. Your chapter-six detail, buried in the middle of a sprawling chat, is exactly the kind of thing that gets skipped.
Long chats get compressed
When a conversation fills up, something has to give. Claude's help center says that as a chat nears the context limit, Claude summarizes earlier messages to make room. A summary keeps the plot and drops the texture: the scar on her left hand, the fact that he lies about his age, the rule that the magic costs blood.
A new chat starts nearly blank
Open a fresh conversation and the model starts over, apart from whatever memory features you've switched on. ChatGPT's memory can carry details between chats, and both ChatGPT and Claude let you attach files to a Project that every chat inside it can read. Those help. But memory carries over what it happens to pick up, which isn't a reliable copy of your manuscript, and a project file only knows what was true when you last uploaded it. Nobody updates it when chapter fourteen changes something.
There's no structure
This is the real root of it. A chat history is a transcript. Nothing in it marks "this is canon" versus "this was a draft I threw away". The version where the brother lived is sitting in the same history as the version where he died, and the model has no reliable way to tell which one you kept.
The manual fixes that actually work
You can get a long book out of a general chatbot. You just have to do the remembering for it. If you're choosing between a chatbot and a tool built for fiction, which AI writing assistant for fiction actually keeps your facts is a ten-minute test.
- Keep a story bible. One document: every character with name, age, look, what they want and what they know; places; the rules of your world; open threads. When canon changes, update it that day. (Novelcrafter calls this a Codex. Same idea.)
- Write a summary after each chapter. A short paragraph: what happened, what changed, who learned what. You can have the AI draft it, but read it and fix it, because a wrong summary is worse than none.
- Keep a running recap. A page of "the story so far", rebuilt every few chapters from the chapter summaries.
- Send a packet, not a history. Start each session, or each chapter, in a fresh chat. Paste the bible, the recap, the last chapter or two, and the beat you want written. A clean packet usually beats a giant chat.
- Mark canon out loud. Tell the model what's decided ("Tomas is dead as of chapter 3, this is final"). Don't leave rejected drafts in the conversation it's working from.
- Run a continuity check. Before you accept a scene, ask the model to list anything in it that contradicts the bible. Models tend to be better at checking against a list than at remembering one. Keep it to one scene at a time: an AI plot hole finder gets much less reliable past about 1,200 words.
This works. It's also a second job. Every chapter you're the writer and the librarian, and the librarian's work is the first thing to slip when the writing gets good.
For the full chapter-by-chapter routine built on this, see how to write a novel with AI without losing the plot.
Tools that do some of it for you
Fiction-specific tools each have their own answer to this, and the big difference between them is who does the bookkeeping.
- Sudowrite has a Story Bible (characters, worldbuilding, outline, synopsis) that you fill in with AI help, and a Saliency Engine that picks which details the AI sees for each request. Chapter Continuity links chapters so the AI treats them as one story, reading up to 20,000 words back. The character cards are generated from your synopsis or written by hand, and you edit them; the docs don't describe them updating from the manuscript.
- Novelcrafter has a Codex that recognises names and aliases as you type and tracks how characters change through progressions tied to a point in the story. You write the entries. Scene summaries are written by you or generated when you run the summarize action. You bring your own AI key.
- NovelAI uses a Memory box and an Author's Note, plus a Lorebook whose entries get inserted into context when their activation keys appear in the recent text. You write and maintain the entries.
In all three, you keep the entries current as the plot moves. KudoWrite does it as you write, which also answers the canon problem above. You write the book a paragraph at a time on a page that is the manuscript; a couple of minutes after you pause, the story bible reads what's new or changed, and the whole-book summary catches up when you leave the chapter. Write the chapter-three beat where her brother dies, and the chapter-seventeen dinner is planned from a story so far in which he's dead.
Each beat is also written from a packet, the way fix 4 says: the bible entries that passage needs, the story so far and the beats on either side. If the bible records a detail wrong, it's in the sidebar to correct by hand.
If your story started as an AI Dungeon adventure, what its memory limits come to in words is worked out on its own page. Bringing a book over: My Books → Import takes .docx, .epub, Scrivener and Novelcrafter projects and most other formats, with story bibles from Sudowrite CSVs, a Novelcrafter codex, a NovelAI lorebook or AI Dungeon story cards, so the entries you already wrote come with it. You get one free credit to try it once your email is confirmed; after that, one-off packs ($3.50 for 100 credits, roughly a short novel; $7.50 for 500; $12.50 for 1,000), with no subscription.
If you're weighing it against a tool you already use, there's a page for each: a Sudowrite alternative, a NovelAI alternative and a Novelcrafter alternative.
Which fix fits your book
- A short story or a few chapters. A good chatbot with a big context window is fine. Keep it in one tidy chat.
- A novel, and you don't mind admin. The manual kit above, in ChatGPT or Claude, works. The bible and the per-chapter summaries do most of the heavy lifting. A serial that runs past a hundred chapters needs the same kit plus a list of open threads; how to write a web novel has a series bible to copy.
- A novel, and you want control. A tool where you curate the memory, such as Novelcrafter's Codex or NovelAI's Lorebook.
- A novel, and you'd rather just write. Something that keeps the bible for you. That's the gap KudoWrite was built for: rewrite the scene your chatbot got wrong in KudoWrite, then read what the bible recorded.