How to keep characters consistent across a long novel

Somewhere around chapter twelve you write that Maren's eyes are grey. You forget that on page nine they were green. A beta reader catches it. You go back to fix it and find that her brother's name has migrated from Tomas to Thomas across three scenes, that the village she grew up in had a river in one flashback and a salt flat in another, and that the locket she gave away in chapter five is somehow back around her neck in chapter eighteen. None of this is laziness. It is what happens when a long novel runs through the soft tissue of working memory.

A 100,000-word manuscript holds more facts than any human can keep correct by recall alone. Every long-novel writer eventually contradicts themselves. The interesting question is not whether it will happen but what catches it before the reader does.

Why the contradictions happen

Two separate failure modes, often confused.

The first is human working memory. You wrote the green-eyes sentence eight months ago. You have written 60,000 words since. You no longer remember that sentence. You write grey because grey is what feels right today. There is no AI involved in this failure. It happens to every novelist who has ever finished a long book without a tracking system.

The second is specific to AI-assisted drafting. Language models do not have persistent memory across sessions. They have a context window, and whatever is in it. If you draft chapter twelve in a fresh conversation without giving the model chapter nine, it will not know about the green eyes. Worse, it will not say "I do not know." It will confidently make something up. That is the part to watch. An LLM without grounding will not admit a gap; it will fill it.

Both failures have the same fix. Stop relying on memory. Move the facts into structured storage that can be queried.

What a character bible actually is

The traditional answer is a Word document. A page per character. Eye colour, hair colour, height, birthday, three childhood memories. Most of it never gets touched again.

That document fails because it is unstructured prose. You cannot ask it questions. You cannot diff it against the manuscript. The AI cannot read selectively from it. By chapter thirty you stop opening it.

A useful character record looks more like a database row. Fields, values, and a clear sense of which fields actually get contradicted in real drafts. The goal is not completeness. The goal is the smallest set of fields that, if tracked, would have caught the last five mistakes you made.

The fields that actually matter

After watching writers contradict themselves repeatedly, the fields worth tracking are smaller than the standard character-sheet template suggests:

• Name and any nicknames, with exact spelling. Tomas and Thomas are not the same.
• Physical details that recur on the page: eye colour, hair colour and length, height relative to other characters, any scar or tell. Not "build" or "complexion." The specific things you have already written into a scene.
• Relationships, stated as pairs. Maren is Joran's sister. Joran is two years older. The pair-form matters because relationships are where most contradictions live.
• Where they were born and where they live now. One line each.
• Their job or function in the story.
• Possessions that have appeared on the page. Especially anything the plot turns on. The locket. The knife. The notebook.
• Any backstory detail that has been stated to the reader. Not the full backstory. The detail you already committed to.

Notice what is missing. Favourite food. Star sign. Middle name. Most character-sheet templates ask for fields no reader will ever see. Skip them. They are not what you will contradict.

What does not belong in the bible

There is a discipline question hiding here. The longer a character bible gets, the less it gets used. Past a certain density, opening the document feels like work, and you stop. A useful character record is read at every scene. An exhaustive one gets read once.

So: leave out the things you can rederive. Leave out the things that have not been written into the manuscript yet. The bible tracks what is canon, not what is potential. A possibility that lives only in your head is not canon and does not belong in the record until you commit it to a page.

The bible has to talk to the prose

This is where the standard advice falls apart. Even if you keep a perfect Word document, it sits in a separate window from your draft. The AI assistant you are using to draft chapter twelve has no idea that document exists.

The fix is structured records that the drafting tool reads from automatically. When you start writing a scene that involves Maren, the system pulls the Maren record into the model's context. When the scene establishes a new fact about her, the system asks whether to write that fact back. The bible stops being a document you maintain on the side and becomes part of the loop.

This is the architecture behind Pendraic's Story Index. Every character, place, object, and rule is a structured record. The drafting layer reads from it before generating prose, and the analysis layer flags when a new scene contradicts what is already known. It is the consistency engine, not a wiki.

Useful even if you are not using AI. Even pen-and-paper writers benefit from a queryable cast list with stable spellings.

A drafting habit that catches most drift

A small ritual goes a long way. Before you start a scene, list every character who appears in it and skim their record. Two minutes. You are not memorising. You are loading.

After you finish a scene, look at what you established about each character. If anything is new (a new detail, a new opinion, a new memory), write it into the record. Two more minutes. The cost is small. The savings, four months later when you are halfway through draft three, are large.

This same rhythm scales to other Registry entries: places, objects, rules. Worldbuilding details drift exactly the same way characters do. The same fix applies.

What to do when you find a contradiction

You will find them. Every writer does. The question is how you handle them.

Resist the urge to fix the early scene. The early version is usually the one your subconscious chose for a reason. If chapter three says her eyes are green and chapter twelve says grey, the rule is: which version does the story benefit from? Pick one, then update the bible, then sweep the manuscript with a search for the wrong version and correct it everywhere.

A character record makes this trivial. Without one, you are doing a manual hunt across a long document with imperfect search terms, and you will miss the seventh instance.

Where Pendraic comes in

Pendraic builds the Registry into the platform as a first-class surface. Characters, places, objects, and rules live as structured records, not as a side document. Every drafting call pulls the relevant records into context automatically. Every new fact the prose establishes is offered as a suggested update to the record, so chapter twelve cannot quietly contradict chapter three without the system noticing.

If you have a half-finished manuscript with the usual drift, sign in and seed the Story Index with what you already have. The act of writing it down catches a surprising amount on its own. The rest gets caught when you draft against it.

Memory is not the writer's tool for a long novel. Storage is.