The Model Doesn’t Know Your Canon Until You Hand It Over
The pitch practically writes itself. Context windows have grown a thousandfold in a few years; the frontier models now take a million tokens and more in a single call, which is comfortably an entire multi-book series. So the continuity problem, the eye colors and timelines and who-knew-what-when that have haunted every long saga since serialized fiction began, should be over. Paste in the whole series. Ask your question. Total recall.
Anyone who has actually tried to run a long series through a long window knows the pitch oversells, and the research now explains why. The window is real. The recall is not uniform. And the writers getting genuine continuity enforcement out of these tools are doing something more deliberate than pasting: they are engineering what the model sees, and the shape of that engineering is worth learning, because it is settling into a consensus.
The Middle of the Window Is a Swamp
The foundational finding, replicated enough times to treat as law: a model’s ability to use information in its context depends on where in the context the information sits. Recall is strong at the beginning of the window, strong at the end, and degrades through the middle, a U-shaped curve that practitioners have taken to calling lost in the middle. Stretch the window to a million tokens and the middle gets longer, which means the swamp gets wider.
There is a second effect stacked on the first. Filling a window does not just risk the model missing a relevant fact; the irrelevant text is not neutral padding. Longer contexts raise the probability that off-target material gets woven into the answer, and measured accuracy on reasoning tasks declines as windows fill with material that merely might matter. The field has a name for this too, context rot, and it should retire the paste-everything workflow on its own.
Map that onto a series. Book one sits at the head of the window, the current chapter at the tail, and the middle books, where the crucial betrayal happened and the timeline got complicated, sit exactly in the swamp. The continuity errors a long window fails to catch are concentrated, structurally, in the part of the story where continuity errors live.
Fine-Tuning Teaches Style, Not Facts
The other tempting shortcut is training: fine-tune a model on the manuscripts so it internalizes the canon. This mistakes what fine-tuning does. Tuning a model on a corpus shifts its voice, its cadence, its defaults, and it does this well; that is the whole art of style transfer. What it does poorly is store facts you can rely on retrieving. A fact learned through tuning is a statistical tendency, and a statistical tendency about your antagonist’s backstory will happily blend with ten thousand other antagonists from pretraining, then assert the blend with total confidence.
Canon is exactly the kind of knowledge that should never live as tendency. It is discrete, exact, checkable, and it changes every time a new chapter lands, which would mean retraining forever. The engineering consensus is blunt about this division of labor: parametric memory for style and capability, external memory for facts. Your canon belongs outside the model, in a store you control, handed to the model at the moment of need.
Retrieve Bounded, Reason Long
Which is why the argument of the last two years, retrieval versus long context, resolved as a merger rather than a victory. The emerging default architecture for serious systems is a hybrid: use retrieval to select a substantial but bounded set of relevant material, then let a long-context model reason carefully over that curated set. Retrieval solves what long windows cannot, keeping the swamp out. Long windows solve what retrieval alone cannot, giving the model room to hold an entire subplot in view while it works.
For a series, the translation is direct. A continuity question about the poison used in book two should pull the poisoning scene, the apothecary chapter, the relevant codex entries, and nothing else, and the model should see those passages in full rather than as fragments.
“Here are the six passages where the poison appears, the apothecary chapter, and the codex entry on the guild’s antidotes. Using only these, check the attached scene for contradictions in timing, symptoms, and who could know the recipe. Cite a passage for every conflict you flag.”
A brief like that is the whole method in miniature. Twenty thousand well-chosen tokens beat a million indiscriminate ones, every time it has been measured. The window’s size is a budget, and budgets reward spending over stuffing.
Narrative Needs Recency, Not Just Similarity
One more finding matters specially for fiction, and it comes from the study of long-running agent systems. Retrieval engines select by semantic similarity, the passages that most resemble the query, and for reference lookups that is the right criterion. For continuing a narrative it is subtly wrong. What a model needs to write the next scene is dominated by what just happened, the last chapters in order, and retrieval by similarity will cheerfully skip them in favor of a thematically similar scene from three books ago. Practitioners running long sessions learned this the hard way: you need the recent history because it is recent, whether or not it resembles the question.
So the working context for drafting has two lanes.
The recent chapters ride along verbatim. In order, always, because narrative momentum lives there and similarity search will not protect it.
The deep canon arrives by retrieval. Selected per scene, as full passages, chosen for the scene being written rather than for resemblance to a query.
Confuse the lanes, retrieve the recent or paste the deep, and you get either a model that forgets what happened yesterday or one lost in the swamp of everything that ever happened.
The Series Bible Becomes a Database
All of this converges on a practical shift in what a series bible is. The traditional bible is prose about prose, a document a human skims. A bible built for retrieval is structured data: entities with attributes, events with dates, relationships with directions, each fact atomic, sourced to a chapter, and tagged for when it becomes true in the timeline. That last field matters more in fiction than anywhere else, because a fact can be true in the world and unrevealed in the story, and a retrieval system with no notion of reveal order will leak a book-five secret into a book-two scene without ever knowing it did something wrong. For anyone drafting early books while later canon exists, spoiler-awareness is the difference between a continuity tool and a continuity hazard.
Structuring canon this way sounds like clerical work, and partly is, but the clerical work pays twice. Once when retrieval starts returning exact facts instead of vibes. And once more when the structure itself starts catching contradictions, because two atomic facts that disagree are visible in a way two paragraphs that disagree never were.
The Discipline Is the Feature
Step back and the pattern across every finding is the same. The tools reward writers who treat the model as a brilliant collaborator with no memory, because that is what it is. It does not know your canon. It never will, in any durable sense, no matter how large the windows grow, and the growth of the windows has turned out to be a reason for more curation rather than less, since a bigger window is mostly a bigger opportunity to bury the relevant under the possible.
The writers holding five-hundred-thousand-word sagas together with these systems are doing it with structure: canon as data, retrieval as librarian, recency as spine, the window as a budget spent deliberately. None of that is glamorous. All of it is learnable. And the alternative, trusting the swamp, has a failure mode with a distinctive signature: confident, fluent, plausible prose that quietly contradicts your own book, which is the one error a long-series writer can least afford, because readers who have followed a world for years are precisely the readers who notice.
