Cost and models
How memory works
What the model can see, and what it looks up.
The hardest problem in long-form AI writing is that models forget. Here is precisely what this one gets.
Recent chapters, in full
Every generation includes the previous chapters as real text, newest first, until a 60,000-token budget is used up. For 2,000-word chapters that is roughly the last twenty chapters verbatim.
Older chapters, as summaries
When you accept a chapter, a summary of it is written and stored — plot events, character developments, revelations, unresolved threads. Once the budget is full, the oldest chapters fall back to those summaries. The novel stops growing in the prompt, which is what stops the cost growing too.
Everything else, on demand
The full text of every chapter stays searchable. When the model is unsure about a name, a rule or a promise made forty chapters ago, it calls a search tool mid-draft, reads the matching passages and continues. You can watch this happen in the context panel — it is the “Searching previous chapters” line.
What this means in practice
- Continuity holds much further than a chat window, because nothing is ever truly dropped.
- Accurate chapter titles help — they are what the summaries and searches key off.
- Editing an accepted chapter rewrites its summary, so later chapters stay consistent with the change.