Developer Spent 5 Months Building AI Coding Assistant Memory After It Failed Basic Recall

A developer discovered that coding assistants cannot recall past project decisions, such as architectural choices or error-handling strategies, when opened in a fresh chat session. After his own assistant failed a simple memory test, he spent five months building a dedicated memory engine for AI coding tools. His key finding was that simple conversation storage and similarity search are insufficient, as they cannot resolve contradictions or match semantically related but differently worded discussions. Instead, he found success by extracting discrete facts from conversations, connecting them to establish which claims supersede others, and constructing answers from those relationships at query time. He also warned that swapping embedding models on an existing database silently corrupts recall without throwing any errors.