The authors propose Hippocam, a hierarchical memory and continual learning architecture for LLM agents, inspired by human memory consolidation. Hippocam structures an agent's work as nested intents, allowing for the consolidation of experiences and the progressive recovery of details.