GraphRAG alone won't fix AI agents that lack a proper memory contract

A software developer argues that poor retrieval in AI agents is rarely a vector index problem — it is a missing memory contract that defines how and where an agent looks up information. Without clear entry rules, even expensive GraphRAG or hybrid retrieval systems will confidently surface wrong results from stale or misclassified sources. The author distinguishes between lookup queries, where vector search typically performs well, and synthesis queries across entity graphs, where GraphRAG has an edge but carries higher maintenance costs. A proposed memory contract covers canonical sources, scoped content categories, hub-first entry patterns, redirect limits, and hard exclusions such as secrets, PII, and legacy backup files. The core diagnostic offered is simple: if an agent cannot identify which hub it entered through to answer a question, the system has an index but no contract.