This research delving into Conversational Memory of LLM Agents underscores the unprecedented investigation into their efficacy over long-term dialogues. The study puts the spotlight on machine-human pipelines, enabling high-quality long-term dialogues that are anchored in personas and temporal event graphs.

Highlights:

• Establishes a machine-human pipeline to produce very long-term dialogues.
• Collects LoCoMo, a dataset featuring extended conversations over multiple sessions.
• Offers a comprehensive evaluation benchmark assessing long-term memory in models.
• Uncovers challenges faced by LLMs in understanding lengthy conversations.

This research opens up new perspectives on the capabilities and limitations of LLMs in long-term interactions, stressing the importance of sophisticated models capable of navigating complex temporal dynamics within dialogues.