LLM agents typically manage conversations within a limited context, but the new LoCoMo dataset pushes boundaries with dialogues spanning 35 sessions with detailed event graphs. Researchers explore the impact of long-context LLMs and RAG on maintaining memory across extended dialogues. Read the full paper.

• LoCoMo dataset includes very long-term dialogues over many sessions.
• Benchmarked tasks for measuring long-term memory in models.
• Highlighted challenges for current LLMs in handling long-term dynamics.
• Use of long-context models and RAG show promise but don’t match human levels yet.

Very long-term memory handling is a leap forward in artificial intelligence, with implications ranging from personal digital assistants to complex interactive applications.