This preprint examines why enterprise AI systems need governed shared memory beyond isolated retrieval-augmented generation workflows. It explains how fragmented context, stale facts, access boundaries, and disconnected AI clients can cause large language model systems to answer from incomplete or outdated organizational knowledge. The article compares long-context prompting, vector retrieval, graph retrieval, and permissioned knowledge graph memory. It argues that enterprise AI systems should preserve source-backed facts, supersession history, conflict signals, team isolation, readonly access, and auditability. It also reports preliminary local benchmarks, including a 100-scenario comparison between vector RAG and a curated knowledge graph baseline, to test accuracy under stale, conflicting, and permission-sensitive information conditions. The accompanying source package includes the article source, synthetic benchmark data, benchmark scripts, Zenodo metadata, and generated LaTeX source for reproducibility.