Long conversations with AI agents are important, but agents often forget previous interactions because their memory is limited. This paper addresses how to help agents remember facts, events, and relationships from long histories.
The authors propose a memory system with two complementary parts: episodic memory that stores events in chronological order, and structured relational memory that extracts facts and connections between entities (people, places). They also add a consolidation process that merges similar memories and removes outdated ones to keep storage efficient.
Combined, the two memory types outperform existing memory frameworks, achieving 90.2% accuracy on LoCoMo and 85.5% on LongMemEval-S. Consolidation reduced memory size by about a sixth with little performance loss. This enables agents to maintain coherent long-term memory without unbounded storage growth.