zy-zmc/tianming-novel-ai-writer

What it solves

It addresses the "memory loss" and inconsistency issues common in AI-generated long-form fiction. Standard AI tools rely on context windows, which lead to forgotten plot points, drifting character personalities, and broken rules as a story grows to thousands of chapters. This system replaces reliance on context with a structured state-management approach.

How it works

Instead of feeding the AI a long history of text, the system treats the story as manageable data. It uses a closed-loop writing process: reading current state $\rightarrow$ AI writing $\rightarrow$ gatekeeping validation $\rightarrow$ final text storage $\rightarrow$ state update $\rightarrow$ next chapter.

Key mechanisms include:

• Generation Gates: Six validation checks (protocol parsing, reference verification, consistency, entity detection, description checks, and blueprint adherence) that a chapter must pass before being saved.
• 15-Dimensional Fact Snapshot: A system that tracks specific variables (character levels, locations, item ownership, plot flags) across chapters. AI must output a CHANGES JSON declaration to update these facts.
• Long-Distance Recall: Uses vector-based semantic search (via local ONNX models) and historical milestones to retrieve relevant plot fragments from thousands of chapters back.
• Humanization: Employs a local MLM model to replace AI-typical phrasing with more natural language while maintaining semantic similarity.

Who it’s for

Web novel authors and AI-assisted writers who need to maintain strict narrative consistency and complex world-building over extremely long series.

Highlights

• State-Driven Writing: Moves beyond context windows to ensure consistency even at chapter 3000+.
• Rigid Validation: Prevents plot holes and character contradictions via a multi-stage gatekeeping system.
• Structured Planning: A four-layer hierarchy (Outline $\rightarrow$ Volume $\rightarrow$ Chapter $\rightarrow$ Blueprint) for detailed story orchestration.
• Local AI Integration: Built-in ONNX models for embedding-based retrieval and text humanization.
• Comprehensive Tracking: Monitors 12 types of changes across 15 dimensions of story facts.