WriteHERE - An open-source AI-powered long-text writing framework that generates extremely long texts in a single pass.
WriteHERE is an open-source AI-powered long-form writing framework developed by a team led by Jürgen Schmidhuber. WriteHERE is based on heterogeneous recursive planning (HRP) technology, dynamically decomposing the writing task into...
What is WriteHERE?
WriteHERE is an open-source AI-powered long-form writing framework developed by a team led by Jürgen Schmidhuber. Based on heterogeneous recursive planning (HRP), WriteHERE dynamically decomposes the writing task into three heterogeneous tasks: retrieval, reasoning, and writing. It manages task dependencies using a directed acyclic graph (DAG) to achieve adaptive execution. WriteHERE can generate professional reports exceeding 40,000 words and 100 pages in a single run, making it suitable for various scenarios such as novel writing and technical report generation. Completely open-source, WriteHERE allows developers to freely utilize heterogeneous agents, potentially redefining the limits of AI writing.
The main functions of WriteHERE
- Generate very long text in one goIt supports generating professional reports of over 40,000 words and 100 pages, meeting complex writing needs.
- Creative and technical content generationGenerate creative stories, novels, technical reports, etc.
- Dynamic information retrievalSearch for relevant information in real time while writing.
- Consistency of styleMaintain a consistent writing style and content coherence.
- Visualizing the writing processThe writing process is illustrated using a task dependency graph.
The technical principle of WriteHERE
- Heterogeneous task decompositionThe writing process is deconstructed into three heterogeneous tasks: Retrieval, Reasoning, and Composition. Each task has a unique information flow pattern; for example, the retrieval task obtains information from external sources, the reasoning task performs logical analysis, and the composition task generates text. Tasks are recursively decomposed into subtasks until they are reduced to directly executable atomic tasks.
- Stateful hierarchical scheduling algorithmTask dependencies are represented using a directed acyclic graph (DAG), with each task having three states: active, suspended, and silent. The system dynamically adjusts the execution order based on the task states to ensure tasks are completed in logical order, supporting real-time feedback and adjustments.
- Mathematical Formal FrameworkThe writing system is abstracted into a quintuple: Agent kernel, internal memory, external database, workspace, and input/output interfaces. The writing planning problem is formally defined mathematically to ensure the executability of each task and the achievement of the final goal.
WriteHERE's project address
- Project official website:http://writehere.site/
- GitHub repository:https://github.com/principia-ai/WriteHERE
- arXiv technical paper:https://arxiv.org/pdf/2503.08275
Application scenarios of WriteHERE
- Novel writingGenerate full-length novels with complete plots and rich characters, supporting creative writing and dynamic plot adjustments.
- Technical ReportWrite structured technical reports, integrating data and logical reasoning.
- Industry AnalysisGenerate professional reports covering industry trends and market analysis.
- academic papersIt assists in writing academic papers, integrating literature, and generating standardized structures.
- Policy documents: Draft policy documents and white papers to generate authoritative and logically sound texts.