Most AI PPT tools on the market have solved the "zero to one" draft problem, then gotten stuck on two things: generation style is one-size-fits-all, so you re-tune it every time; and one change means redoing the whole slide, so multi-round iteration mostly fails. MemSlides was born for exactly these two pain points — it is a memory-driven Slides agent framework that topped the HuggingFace leaderboard, proposed jointly by Tsinghua University, Shanghai Jiao Tong University, and Beijing University of Posts and Telecommunications, and it is open source.

What MemSlides Is

MemSlides makes "personalization" and "multi-round revision" first-class citizens of a PPT agent. The key is not flashier template generation but remembering how you make slides.

How it differs from general AI PPT tools:

• General tools: every conversation starts from zero. You say "blue corporate style," it generates a version; next time you have to say it all over again
• MemSlides: it stores your preferences (palette, font sizes, chart styling, verbal revision habits) in a memory module and applies them automatically on later generations, and it supports "change only this element on this page"

MemSlides builds "personalized customization" and "multi-round local editing" in as framework-level capabilities, rather than improvising them with prompt engineering.

Core Capabilities

Personalized style memory

The first time you use it, you describe your style preferences just once:

"My PPT preferences: dark backgrounds, sans-serif fonts, Tableau-style chart colors, no more than 3 bullets per page."

MemSlides stores these preferences in structured form in the memory module. From then on, whatever topic you generate slides for, they apply automatically — no repeating yourself every session.

Memory accumulates: mention in passing "add data labels to charts from now on," and the next generation picks it up automatically.

Multi-round local edits

This is MemSlides' core difference from its peers. It supports iterative changes targeting a single page or a single element without rebuilding the whole document.

A typical conversation flow:

1. You: "On page 5, change the pie chart to a bar chart, sorted by revenue descending."
2. MemSlides locates the chart object on page 5
3. It regenerates that element locally, leaving every other page untouched
4. It updates memory (you prefer bar charts sorted by value descending)

By contrast, traditional AI PPT tools regenerate the entire deck on any edit, and earlier fine-tuning is often lost.

Generation quality

MemSlides topped the relevant HuggingFace PPT evaluation leaderboard mainly on two strengths:

• Structural soundness: outline hierarchy and section transitions match presentation logic better than general tools
• Visual consistency: unified styling across pages — no "each slide looks made by a different person" problem

How to Use It

Preparation

• Project location: follow the paper / HuggingFace repository (searching MemSlides will get you there)
• Environment: Python 3.10+, GPU inference recommended
• Dependencies: PyTorch, transformers, and the slide rendering library the paper specifies (usually based on python-pptx or reveal.js)

The Typical Three-Step Flow

Step 1: Initialize your "user profile"

On first run, give MemSlides a preference description. Cover:

• Visual: palette, fonts, font sizes, chart style
• Structure: bullets per page, section-splitting preferences
• Tone: technical briefing / pitch deck / teaching

The more specific this description, the higher the quality of later generations.

Step 2: Generate the first deck

Enter a topic and your key points, for example:

"Generate a 15-page deck titled 'Our SaaS Product Q3 Review,' covering GMV, retention, churn analysis, and next quarter's plan."

MemSlides pulls from the memory module to apply your preferences and generates the first version.

Step 3: Iterate with local edits

Change the pages you're not happy with:

• "Change the line chart on page 3 to a bar chart"
• "Page 7 has too many bullets; merge them into 2"
• "Make the overall palette darker"

Every edit also updates memory, so the next generation applies it automatically.

Tip: during multi-round editing, change one dimension at a time (content first, then style) so problems are easier to localize.

Best-Fit Scenarios

Comparison With Peer Tools

Common Issues

• What if the memory module "misremembers": MemSlides lets you view and edit memory entries; if you spot a wrong preference, just delete that entry.
• Can preferences be shared across a team: the current version targets a single user; team sharing means manually exporting/importing the memory config.
• Generation is slow: multi-round edits are much faster than the first generation (only local recomputation); run the first generation on a GPU.

---

Related Links

• BAAI Community coverage: https://hub.baai.ac.cn/view/56268
• HuggingFace model page (search MemSlides)