This research proposes a novel system for generating long-form narratives, such as novels or movie scripts, by integrating techniques from LongWriter and the AI-agent town concept. We aim to develop a multi-agent system with extended context memory, enabling the collaborative creation of coherent, engaging stories exceeding 10,000 words.
Introduction and Objectives
Current large language models (LLMs) struggle with generating high-quality long-form content due to limited context windows and challenges in maintaining long-term coherence. This project addresses these issues by combining multi-agent interactions with extended memory techniques. Key objectives include developing a collaborative story generation system, implementing extended context memory, and evaluating narrative quality and coherence.
Methodology
Multi-Agent System Design
We propose creating AI agents with distinct roles, personalities, and backstories, such as characters, narrators, and plot managers, capable of inter-agent communication.
Extended Context Memory
Using adapted LongWriter techniques, we aim to implement hierarchical memory structures for each agent, including short-term, long-term, and thematic memories.
Narrative Generation Framework
A modular system will be designed for generating scenes, managing plot progression, and resolving conflicts between agent actions. A central plot management agent will guide overall narrative structure.
Implementation and Evaluation
We will utilize pretrained LLMs like GPT-4o as a baseline, incorporating custom attention mechanisms for long-range dependencies and distributed computing for multi-agent operation. Narrative coherence, character consistency, and thematic development will be evaluated through automated metrics and human assessments, comparing outputs with human-authored content.
So, the multi-agent architecture is intriguing, but the specifics of inter-agent communication need clarification. Will communication be entirely emergent, a sort of free-for-all negotiation amongst agents with potentially unpredictable outcomes? Or will a structured protocol, perhaps incorporating explicit communication constraints, be implemented to ensure narrative coherence?
Wow, would be really cool to explore the difference in output between 2 systems. First system would involve agents primed to think they are the actual characters in the plot of the story. Second system would involve agents primed to think they are the writers of a story (reflecting the reality).
I wonder how the stories would turn out differently between these kinds of situational setups?
Agree on more details required on the inter-agent interaction process.
Perhaps the best approach would be a hybrid of the options you identify Muddy. For example, using one agent as a Project Manager and this agent facilitates the structured development of the story. Within this agent’s prompting of the other agents, they are given total creative freedom to come up with the specific story details; however, the Project Manager grants the other agents this emergent communication only within the framework of its story development.
e.g. “Okay, as characters in this story, illustrate what happens in Scene 1” —> and it is from here that the magic of agents having emergent communication is realized
Overall, I think designing a protocol with options will be most effective. It will likely be easiest to facilitate coherent, lengthy stories with the approach I mentioned above, however users should still have the option to dramatically switch the story structure with the help of AI, and the protocol should adapt to this.