Why ChatGPT Forgets Your Characters (and How to Finish the Novel Anyway)

Why ChatGPT Forgets Your Characters (and How to Finish the Novel Anyway)

Chapter one was wonderful. By chapter nine your detective has a different eye color, her dead mother is phoning her, and the villain has become a reasonable man who just wants to talk. You did not change anything. ChatGPT did.

This is one of the complaints that comes back again and again in our support chat from people who tried to write a novel with a chatbot. The good news is that the causes are well understood, and once you know them you can work around most of it.

We make BookNova, an AI novel writer, so we have an interest in this topic. The ChatGPT limits below come from OpenAI's pricing and help pages, read on November 22, 2026. The research is linked so you can check it.

Short answer

  • Every chat has a fixed window. On the Free plan the everyday model holds 27,000 tokens, about 20,000 words, including its own replies.
  • When the window fills, details go. A summary may keep the plot and lose the scar, the stutter and your rules.
  • Memory is not a manuscript. OpenAI says it does not retain every detail.
  • Fix: a one-page canon sheet, one chat per chapter, a state update after each chapter, and an audit before you move on.

Reason 1: the window is smaller than your book

A language model has no memory between messages. Each time you press Enter, ChatGPT sends it the conversation again, as much as fits. That amount is called the context window.

OpenAI publishes it on the pricing page. Here it is in words, at roughly three quarters of a word per token:

PlanEveryday model (Instant)In wordsReasoning modelsIn words
Free27K tokensabout 20,000VariesVaries
Go54Kabout 40,000256Kabout 190,000
Plus54Kabout 40,000256Kabout 190,000
Pro128Kabout 96,000400Kabout 300,000

Everything counts: your notes, your instructions, every draft it wrote and every "make it darker" you typed. A chapter of 2,500 words that you regenerate twice takes 7,500 words of the window. On the Free plan you can run out during chapter three, and at best around chapter eight.

Notice the second half of the table. On Go and Plus the reasoning models have almost five times the room of the everyday model. If you are paying and still writing with Instant, switching is the simplest improvement you can make.

We measured it

In November 2026 we had GPT-6 Luna, the model behind ChatGPT's free plan, write ten chapters of about 3,200 words in one conversation through the API, where the window is far larger. We recorded how big the conversation was before each chapter.

Before chapterTokens in the conversationFree plan window (27K)
37,771Fine
516,498Filling
621,385Nearly full
726,400Full
831,999Over
1040,158Over

That is with no retries and no discussion. Any of both, and the story bible falls out of a free chat around chapter five.

And with everything still in view? The model kept the facts we checked: the scar stayed on the left hand, the brother stayed deaf in the right ear and the fire stayed nine years ago, through all ten chapters. So did GPT-6 Sol and Claude Sonnet 5.5. Gemini 4 Neon broke one rule. The model is not forgetful. The window is.

Reason 2: what happens when it fills

OpenAI's help pages describe a "Context window full" notice, and sometimes an offer to continue in a new chat with a summary. They are open about the cost. As quoted by SSD Nodes, OpenAI warns that a summary "may leave out an earlier instruction or the exact wording of something you wrote."

Think about what that means for fiction. A summary keeps events: Mara found the ring, Tomas came home. It drops texture: that Tomas stutters only when he lies, that nobody in town says the father's name. Those small facts are the characters.

Reason 3: even a full window is not attention

A model can also have chapter three in view and still contradict it, above all in longer and busier books than our test.

In 2026 a research team published ConStory-Bench, which tested a range of models on stories of 8,000 to 10,000 words, short enough to fit in most paid windows. They found that:

  • Errors grew roughly in line with length. Twice the story, about twice the contradictions.
  • The most common kinds were facts and timeline: who is where, what is called what, how many days passed.
  • The facts were usually set up 15 to 30 percent of the way in, and contradicted at around 40 to 60 percent.

So the saggy middle is real, and it is measurable. It is also why buying a bigger plan helps but does not cure it.

Why characters turn nice

Forgetting also has a moral direction. Assistants are trained to be helpful and agreeable, and when your character sheet drops out of view, the character slides toward that default. Bitter people apologize. Liars explain themselves. Enemies talk it through.

If your villain has become pleasant, it is often a sign that the description left the window several chapters ago.

Why Memory does not solve it

ChatGPT's Memory is for preferences and personal context. OpenAI's help page says it "does not retain every detail from every conversation" and that ChatGPT decides what is relevant. That suits "I write dark romance in past tense." It does not suit 214 facts about a town.

Projects come closer. They keep chats, files and instructions together: up to 5 files on Free, 25 on Go and Plus, 40 on Pro, according to the Projects article. Uploading your story bible there is worth doing. Just remember that uploaded files are consulted, not memorized word for word.

A routine that works

1. Write a canon sheet of one page

Not a wiki. One page you can paste. For each main character: looks, want, wound, secret, voice, and one thing they would never do. Then the rules of the world and the timeline.

info Canon sheet template

NAME, age, job. LOOKS: two facts only. WANTS: ... AFRAID OF: ... SECRET: ... KNOWN BY: ... VOICE: short sentences, never swears. NEVER: apologizes first. CHANGES AT: chapter 12.

2. Start a new chat for each chapter

Paste three things: the canon sheet, a half-page summary of the story so far, and the last 300 words of the previous chapter. That is about 1,500 words and leaves nearly all the window for the work.

3. Give scenes, not chapters

Ask for one scene of 800 to 1,200 words with a goal, a conflict and an ending. Short requests follow instructions better and waste less when you reject them.

4. Update the state before you close the chat

Use this prompt at the end of each chapter:

info State update prompt

List every fact this chapter established or changed. For each main character: where they are, what they now know, injuries, promises, lies told. Then list open questions the reader is waiting on. Facts only, no commentary.

Paste the answer into your own document. That document is the book's memory from now on.

5. Audit every five chapters

Open a fresh chat with a reasoning model, paste the canon sheet and five chapters, and ask: "List contradictions only. Quote both passages." Asking for quotes matters, because it stops the model from inventing problems.

6. Protect the hard edges

Repeat in each prompt what the character must not do. "Ines does not forgive her in this scene" works better than a paragraph of psychology.

How much work is that?

An honest estimate for an 80,000-word novel in 32 chapters: 32 new chats, 32 state updates, about 100 scene prompts and six audits, plus reading everything. Plenty of people finish this way. The writing is fast. The bookkeeping is the job.

Let software do the bookkeeping

If you like writing with AI but not being the archivist, use a tool that keeps the state outside the chat.

  • Novelcrafter has the Codex, a story wiki that travels with your prompts. You bring your own AI key.
  • Sudowrite has the Story Bible and linked chapters.
  • BookNova plans the novel with you and writes the chapters in order while tracking characters, what they know and which threads are open. You can steer each chapter or let it run. If you already have chapters from ChatGPT, the Platinum plan imports a .docx and continues. It is a single payment.

None of these is immune to mistakes, ours included. Read the book before you publish it.

Bottom line: ChatGPT forgets because it was built for conversations, and a novel is not one. With a canon sheet, a chat per chapter and a state update it becomes a good drafting partner. If that sounds like a second job, hand the memory to a tool made for it.

FAQ

Does ChatGPT Plus remember a whole novel?

With reasoning models Plus has a 256K window, about 190,000 words, so a typical novel fits. With the everyday model it has 54K, about 40,000 words. Fitting is not a guarantee of consistency.

Does deleting old chats help?

No. The window belongs to the current conversation. Deleting other chats or memories does not free space in it.

Does a custom GPT remember my book?

It keeps its instructions and files, which helps. It has the same window limits as the model it uses.

Do Claude and Gemini forget too?

Yes, but their windows differ. Paid Claude plans go up to 1M tokens, Gemini has 32K on its free plan and 1M on AI Pro. See ChatGPT vs Claude vs Gemini.

Sources, all read November 22, 2026

BookNova AI Assistant

Ask me anything about our platform