huohua325/memslides

MemSlides is a research project and tool that generates presentation slide decks with the help of an AI agent, and treats making slides as an ongoing back and forth process rather than a single, one time conversion of source material into a deck. Instead of regenerating the whole presentation every time you ask for a change, it tries to remember context and only rewrite the specific slide or slide area you asked to change. The project is built around three kinds of memory that work together. Long-term user profile memory remembers your recurring preferences, such as your favorite visual style or layout choices, across different presentation projects over time. Working memory keeps track of temporary preferences and constraints during a single editing session, so instructions you give partway through revising a deck stay consistent as you continue to give feedback. Tool memory stores past experience about which editing steps worked reliably, so future edits of a similar kind are less likely to fail or need repeating. When you give feedback on a deck, MemSlides tries to figure out the smallest part of the presentation actually affected by that feedback and updates only that part, rather than regenerating every slide from scratch, which helps keep the rest of the deck stable and unchanged. The project accompanies a research paper and has an online demo at memslides.com. Running it yourself requires installing several system dependencies including LibreOffice and font packages, setting up a Python environment through Conda, installing a browser automation tool called Playwright, and providing your own AI model credentials, which are kept outside the code through environment variables or a private configuration file. A Docker based setup is also available as an alternative to installing everything directly on your machine. The project is released under the Apache 2.0 license, which allows free use, modification, and commercial use. MemSlides is aimed at AI researchers and developers interested in building or studying agents that generate and revise presentations, rather than at people simply looking for an everyday slide making tool.