Stop regenerating the whole deck for one typo

What it does MemSlides is a Python framework that turns presentation generation from a one-shot conversion task into a stateful authoring process. It maintains three tiers of memory—persistent user profiles, active session working memory, and reusable tool memory—to drive personalized, multi-turn slide creation and revision. When you ask for a change, it projects feedback onto the smallest affected slide region rather than regenerating the full deck.

The interesting bit Most LLM slide tools treat each prompt as a blank slate; MemSlides treats your presentation history as a first-class citizen. The intent-conditioned user profile memory is the unusual piece—it routes preferences by presentation intent, so the agent knows you prefer minimal titles for investor pitches but detailed ones for research reviews.

Key highlights

• Hierarchical memory stack: long-term user profiles, temporary session constraints, and cached tool execution experience.
• Scoped local revision patches only the affected slide region instead of rewriting the entire deck on every feedback turn.
• Intent-conditioned personalization that routes preferences by presentation type (theme, layout, content strategy).
• Runs in Docker for reproducible research experiments, with a live demo at memslides.com and an accompanying arXiv paper.
• Apache 2.0 licensed.

Caveats

• Requires user-provided model and service credentials; it is an orchestration layer around external generators rather than a self-contained slide engine.
• The built-in verification suite is minimal, and real generation experiments demand custom configurations and private credential files.

Verdict Researchers and developers building stateful agent pipelines should look here. Skip it if you need a drop-in slide maker that works without API keys and custom configs.