MemSlides: PPT Agent with Memory

量子位 · wechat · 2026-07-10

Researchers from Tsinghua, SJTU, BUPT, and other institutions proposed MemSlides, a memory-driven SlidesAgent framework designed for personalized slide generation and multi-turn local edits. By hierarchically modeling long-term preferences, current task constraints, and tool usage experiences, it solves the common AIPPT issues of "forgetting context, poor edits, and accidental deletions."

The article details the framework design, memory mechanisms, and experimental results: a 96.3% closed-loop completion rate, a 53.4% strict pass rate, and the time-to-first-correct-edit dropping to 242.5 seconds. The related paper is published on arXiv and has gained significant traction on HuggingFaceDailyPapers and GitHub.

Original post →

More from coding & agent

coding & agent channel →