Memory Systems

This walks you through building memory architectures for agents that need to remember things across sessions. It covers the progression from simple vector stores to knowledge graphs to temporal knowledge graphs, explaining why vector similarity search falls short when you need to reason about relationships. The benchmark data is useful: Zep's temporal KG hits 94.8% accuracy on the DMR benchmark while GraphRAG shows 20-35% gains over baseline RAG. The layered approach is practical, breaking memory into working, short-term, long-term, entity, and temporal layers. If you're building anything that needs to track entities or relationships over time rather than just retrieving similar documents, the graph-based sections will save you from reinventing this yourself.