Memory Systems

This walks you through memory architecture for agents that need to persist state beyond a single session. It covers the spectrum from simple context windows to vector stores, knowledge graphs, and temporal graphs, with guidance on which frameworks (Mem0, Zep, Letta, Cognee) fit which use cases. The most useful bit is the benchmark insight: Letta's filesystem agents hit 74% on LoCoMo with basic file ops, beating Mem0's specialized tooling at 68.5%, which suggests you should start simple and only add graph structure when you actually need multi-hop reasoning. Good reference when you're deciding how much memory infrastructure your agent really needs versus what sounds impressive.