sigpanic/goink

What it solves

Goink is a desktop AI writing system designed for long-form fiction writing. It solves the problem of "memory loss" in general-purpose LLMs, where the AI forgets character names, plot points, and foreshadowing across dozens of chapters. It replaces the manual effort of reminding the AI about thestory's state and maintaining consistency across a massive text body.

How it works

Goink uses an Agent-based architecture rather than a simple pipeline. An AI agent with access to 31 structured tools allows the LLM to autonomously decide how to query character profiles, update plot arcs, and manage world-building data.

Key technical components include:

• Structured Memory: Dedicated tracking for character relationships (as directed graphs), foreshadowing status, narrative arcs, and reader knowledge.
• Local Semantic Search: A local engine using BGE Chinese semantic models (ONNX Runtime) and sqlite-vec for vector indexing, allowing the AI to find relevant passages by meaning rather than keywords.
• Skill System: A modular system where writing methodologies are defined in Markdown files, allowing users to apply specific styles or workflows via three layers of override (Novel > User > Built-in).
• Diff-based Editing: To prevent the AI from overwriting content blindly, it generates a Diff preview that the user must approve before changes are committed to the text.
• Version Control: Integrated Git for automatic commits and easy rollbacks.

Who it’s for

Fiction writers and novelists who need an AI assistant that maintains strict narrative consistency and structural integrity over long-term projects.

Highlights

• Agentic Maintenance: Automatically triggers checks for character updates and plot arc progression after each chapter.
• Style Distillation: Can analyze a sample text to extract a writer's style (sentence structure, tone, etc.) and generate a custom "imitation skill."
• Visualized State: Provides visual maps for story arcs and location graphs.
• Zero-Config Local Search: Local vector search runs entirely on the machine without needing a GPU or external API.
• Sandboxed Security: Uses a whitelist and SafePath to ensure the AI cannot access or modify unauthorized files.