Towards Usable, Privacy Respecting Long-Term Memory for LLM-based Conversational Agents

In recent years, Large Language Models (LLMs) have evolved from stateless text generators into conversational agents (CAs) that use long-term memory to facilitate personalised and compelling interactions. While these emerging capabilities improve user experience, they also introduce ethical and privacy concerns. Although prior work has examined LLM privacy from a technical perspective, few studies have designed features or interaction techniques for supporting users with privacy management. The aim of my thesis is to address this gap by co-designing usable, privacy-respecting memory mechanisms for LLM-based CAs. I present a mixed-methods survey of LLM-based CA users’ privacy attitudes and strategies, and an ongoing design taxonomy of conversational privacy controls. I then present a plan for two further studies focused on co-designing and prototyping interface controls with users. I aim for my thesis to contribute tangible mechanisms for enhancing human autonomy in memory-enabled AI interactions.