Paper Club: Memory, Modularity, and Safety: Memento as a Case Study in Safe and Performant AI Systems
- Date
- 2025-10-17
- Host
- Singapore AI Safety Hub (SASH) Events
About this event
Modern AI systems do not fail only because models are wrong. They also fail because memory grows messy, components stop composing cleanly, and safety checks become an afterthought instead of part of the architecture. This Paper Club session uses Memento as a concrete case study to explore what it really takes to build AI systems that are both high-performing and safe in practice. About the Event This is an in-person Paper Club meetup focused on reading, discussing, and unpacking ideas at the intersection of memory, modularity, and safety in AI systems. Rather than staying at the level of broad claims, the session centers on a specific case study so the conversation can stay technical, grounded, and useful. The goal is to create a space where people can think carefully about system design choices that matter for real-world autonomous and AI-powered applications. If you care about how agents store context, how components interact, or how safety considerations shape architecture decisions, this discussion is designed to give you material worth wrestling with. Paper Club is also a community format, not a one-way lecture. You should expect discussion, interpretation, comparison of viewpoints, and practical thinking about how the ideas in the paper connect to current AI workflows and system design patterns. What to Expect The session will likely move through the paper in a way that helps both close readers and first-time participants stay engaged. Expect a structured conversation around the paper's core claims, design choices, and implications for building safer and more capable AI systems. You can expect the meetup to include: A framing introduction to the paper topic and why Memento is an interesting case study A guided discussion of the paper's main concepts, especially around memory and modular system design Focused conversation on safety, including how architectural choices can support or undermine reliability Open Q&A and peer discussion where attendees can challenge assumptions, connect the ideas to their own work, and compare interpretations Networking time with others interested in AI, autonomy, and research-driven system design This is not just about summarizing a paper. The value comes from testing the ideas: where they seem strong, where they create tradeoffs, and how they might apply across different kinds of AI systems. If you are newer to paper-reading groups, the format should still feel approachable. You do not need to arrive with a polished opinion. Curiosity, thoughtful questions, and a willingness to discuss architecture and safety tradeoffs are enough to participate meaningfully. Why Attend AI conversations often jump straight to model capability, but many of the hardest engineering and safety problems live at the systems layer. This event gives you a chance to look directly at that layer through a case study that can anchor a serious, practical discussion. You will leave with a better sense of how memory affects behavior over time, how modularity changes the maintainability and controllability of AI systems, and why safety cannot be separated from system design. For builders, researchers, and technically curious community members, that combination makes this a high-value topic. Attending can be especially useful if you want to: sharpen how you read and evaluate AI systems papers understand tradeoffs beyond headline performance metrics discuss agent and autonomy design with people who care about implementation details build shared language around reliability, control, and architecture meet others interested in thoughtful, technically grounded AI conversation Because the event is in person, there is also a clear community benefit. Discussions about architecture and safety tend to get better when people can probe ideas in real time, push on edge cases, and continue the conversation after the formal discussion wraps. Practical Details When: Friday, October 17 at 4:00 PM GMT+8 Location: In person This is an in-person meetup, so plan to attend on site and leave enough buffer to arrive comfortably before the discussion begins. Since the event is built around shared conversation, being present from the start will help you follow the framing and participate more fully in later discussion. The event sits at the intersection of AI, autonomy, community, networking, and meetup culture, so expect a room with a mix of research-minded and builder-minded attendees. That blend is part of the appeal: the conversation should be rich enough for technical participants while still open to broader perspectives on how safe and performant AI systems should be designed. If this topic is close to your work, your research interests, or the questions you keep running into when thinking about agentic systems, this Paper Club is a strong reason to show up in person and think through the ideas with others.
Who should attend
This is for people who want a more serious, grounded conversation about AI systems than the usual surface-level discussion. - You build or study **AI agents, autonomous systems, or LLM-powered workflows** and want to think more carefully about memory, architecture, and control. - You are a **researcher, engineer, student, or technical generalist** who enjoys learning through papers and discussing ideas with others in real time. - You care about **AI safety in practical system terms**, not just abstract principles, and want to explore how safety shows up in actual design decisions. - You are interested in **modular systems design** and want to understand how composability, maintainability, and reliability connect. - You like meetups where you can both **learn something concrete** and **meet thoughtful people** working on related problems. - You do not need to be an expert on Memento already; this is also a good fit if you are simply curious, ask strong questions, and want to deepen your understanding of safe and performant AI systems.