How to keep an AI from forgetting your characters
Continuity is the hardest part of writing a long book with AI. Here's why it breaks and how to fix it.
5 min read
Why AI forgets
Large language models only 'remember' what's inside their context window. As your book grows past a few thousand words, earlier chapters fall out of context and the model starts inventing details that contradict what you already wrote.
This is why a protagonist's job, a sidekick's accent, or a city's layout quietly mutates halfway through an AI-written novel.
The fix: an external, structured memory
The reliable solution is to store story facts outside the model and feed the relevant ones back into each chapter. Characters, places, relationships, and unresolved threads live in a structured record, not in the chat history.
InkSmith calls this the Gita layer. A dedicated agent (Smriti) reads each finished chapter and updates the record, so the next chapter is written against an accurate, current memory of the whole book.
What good memory enables
When memory is solid, the book reads like one author with one voice instead of ten loosely related short stories. Callbacks land, foreshadowing pays off, and characters stay themselves.
It also makes editing safer: change a fact once and the system can keep later chapters consistent with it.
Ready to write yours? InkSmith is free to start.