DraMaze launches multi‑agent pipeline with future‑aware memory for diverse microdrama scripts
agents cohere
| Source: HF Papers | Original article
Researchers introduce DraMaze, a multi‑agent pipeline with future‑aware memory that improves diversity and long‑range coherence in microdrama script generation.
Fudan University and Tencent have unveiled **DraMaze**, a multi‑agent pipeline designed to generate long‑form microdrama scripts with both narrative diversity and coherence. The system tackles two persistent hurdles in AI‑driven scriptwriting: the tendency of candidate continuations to converge on a narrow set of plotlines and the difficulty of maintaining logical consistency over extended scenes.
DraMaze addresses these issues through a combination of trope‑memory‑guided evolutionary tree search and a “Backward Narrative Dependency Tree.” The former steers a population of agents toward varied causal paths by mutating and recombining story elements stored in a curated trope memory. The latter organises past narrative fragments in relation to future plot goals, enabling the agents to retrieve and reuse relevant memories as the story unfolds. This future‑aware memory architecture reduces repetitive content by **32.1 %** and lifts logic evaluation scores by **8.5 points**, according to the authors’ internal benchmarks.
The launch arrives as the microdrama format—short, episodic video narratives popular on social platforms—continues to surge, driving demand for scalable, high‑quality script production. By delivering richer plot variation without sacrificing long‑range coherence, DraMaze could lower the creative bottleneck for creators and platforms that rely on rapid content turnover.
Looking ahead, the research community will likely watch how DraMaze integrates with downstream stages such as storyboard generation and visual synthesis, areas already explored in related multi‑agent frameworks. Adoption by content studios or integration into existing AI‑creative toolchains would provide real‑world validation, while further refinements to the memory‑guided search could push the boundaries of automated storytelling even further.
Sources
Back to AIPULSEN