A benchmarked comparison of three memory architectures for multi-agent LLM systems: raw history dump, vector-only RAG (TF-IDF), and a context graph using NetworkX. The core finding is that vector RAG fails on 'join queries' — questions requiring two separately-stated facts to be combined — scoring only 20% accuracy versus the context graph's 80%. The context graph also achieves 88.9% overall accuracy versus 61.1% for raw history and 50% for vector RAG, while using 18x fewer tokens per query than the raw dump. The post details two real bugs discovered during development (entity vocabulary mismatch and stale-fact retrieval), explains the deterministic benchmark methodology with zero LLM calls, and discusses production considerations including entity extraction costs, alias tables, and fact supersession handling.
