Ask a writer who has spent a year building a novel inside ChatGPT where their story is, and the honest answer is: everywhere and nowhere. The characters exist. The world has rules. The plot has turned a dozen times and found its shape. All of it is in the chats - and that is the problem, because a chat is a transcript, not a story, and a transcript of a year's thinking is close to unusable as a place to build from. This guide is about getting the story back out: what actually makes it hard, and the order of operations that makes it tractable.
The pile at the end of a year
The people who hit this hardest are not the dabblers. They are the ones who took it seriously - who worked a whole book through the chat, refining characters over dozens of sessions, arguing themselves in and out of plot decisions, building a world detailed enough to contradict itself. What they end up holding is not a manuscript and not an outline. It is a mass of conversations, sometimes millions of words across a hundred threads, that contains a real novel and gives you no way to lift it out.
The advice that circulates for this is a coping strategy dressed as a method: split the export into hundred-thousand-word chunks, paste them into another model, ask it to pull out the important bits, and hope nothing goes missing. It is laborious, it is lossy, and the person doing it already knows they cannot trust the result - which is why they keep putting it off. The volume is not the real obstacle, though. The obstacle is that everything valuable in there has no shape.
A chat has no place to keep anything
Here is the diagnosis worth sitting with. In a conversation, a character is not an object you can point to. She is forty separate mentions, spread across weeks, some of them contradicting each other, none of them authoritative. The same is true of every rule in your magic system and every beat in your plot. Nothing in a chat is stored; it is only said, again and again, in slightly different words.
That is why the drift feels unfixable. When the model "forgets" who someone is or quietly rewrites a rule, it is not failing at memory in a way a sharper prompt would cure. There was never a canonical record for it to be faithful to. A story that holds together works the opposite way: each character, place and rule lives in exactly one place, holds its current state, and everything else refers back to it. That single home is what keeps continuity from slipping as a book grows - and it is precisely what a chat does not have. Recovering your story means giving each part of it a home, which the conversation never provided.
What you kept, and what you only tried
There is a second reason the export resists a simple extraction, and it is the one automated approaches miss entirely. A year of working a book leaves the same thing sitting in the chat in several different states, and they are not all canon.
Some of those states are you changing your mind. You called her Mira for three months, then renamed her Wren. You made the mentor the traitor, then decided it was the sister. Here, only the last version counts; the earlier ones are scrap, and keeping them would poison the canon with decisions you already reversed.
Other states are the story itself moving. She begins the book timid and ends it fierce. The kingdom stands in chapter one and has fallen by the finale. The secret is buried, then surfaces. That is not you contradicting yourself - it is the arc, and the whole point of it is that the state changes. Flatten it to a single snapshot and you have thrown away the thing that makes it a story.
A find-and-replace cannot tell these two apart. They look identical on the page: one thing, two values. Only judgement can separate a reversal you want dropped from a trajectory you want preserved - which is the real reason this work cannot be handed to a script and walked away from.
The model already understands your story
The instinct, faced with all this, is to do the sorting yourself, by reading. It is the wrong instinct, and it is why the project stalls. Reading a year of your own conversations to reconstruct your own decisions is enormous, dispiriting work, and you are not even the party best equipped to do it.
The model that was in the conversation is. Telling what you kept from what you tried, and a reversal from an arc, is a comprehension task - reading a long, messy, human exchange and understanding what the person landed on. That is the thing language models are genuinely good at, far more than you would be on your fourth hour of re-reading chat logs. The intelligence to do this was never missing. What was missing was somewhere to put the answer.
Give the model a destination - a structured place where a character becomes a record, a rule becomes a document, an arc becomes a tracked thread - and the excavation turns into something it does alongside you. It reads a conversation, proposes what the canon in it actually is, and writes each piece into its home, merging the forty mentions of your protagonist into one entity even when you spelled her name three ways and never once called her by surname. Connecting the AI you already write with to a canon store that can hold all this is what makes that possible; without a destination, asking a chatbot to organise your chat only ever produces another chat.
Where the canon actually goes
It helps to be concrete about what "structured" means, because it is not a folder of notes. Your characters become entities, each with its own bible that records not only who they are but how they change across the arc - the start state, the turn, the end state - so the movement you were careful to preserve is captured, not flattened. Your world - its places, factions, rules and systems - becomes a set of reference documents the writing can draw on. And the plot's through-lines become the arc itself, laid down first, because everything else is positioned against it.
Notice what this is and is not. You are recovering the canon - the bible - not rewriting the book. The chapters come afterwards, built from clean canon as beats you shape and order, rather than conjured out of the chat transcript. That separation is deliberate: get the foundation solid and true first, and the drafting has something dependable to stand on. It is also what keeps the finished book feeling like yours - the decisions stay recorded as decisions you made, not buried in a log you can no longer navigate.
Getting it out
So the order of operations is short, and it inverts the re-read-everything approach. Take stock first: separate the export losslessly into its individual conversations - a purely mechanical step, provably faithful, with nothing lost or mixed up - and get an inventory of what is genuinely in there. Then qualify: go through that inventory and rule on what is canon and what is scrap, and in what order to capture it. You decide; the model proposes. Only then capture: lay down the arc, then position the cast and world along it, one entity per character, the moving states recorded as the trajectory they are.
The reason to do this inside the AI you already write with, rather than as a filing project on the side, is that it keeps you in the conversation instead of starting over. This is the migration Novelmint is built to take: connect it to your AI client, point the model at your chats, and it works through them with you - taking stock, letting you qualify, and writing the canon it finds into a structure that finally holds it. A year of story stops being a pile you dread opening and becomes a foundation you can build the actual book on.