Designing Persistent Intelligence in Marketing Agents: A Hybrid Memory–Retrieval Framework for Context-Aware Knowledge Generation

Abstract

Large language models that enable enterprise marketing agents must have both correct knowledge underpinning and ongoing contextualization to provide quality and personalized experiences. Traditional Retrieval-Augmented Generation (RAG)is an effective system that enhances the factual accuracy, by using external sources of knowledge, dynamically; but the system is stateless in nature and it does not provide continuity between user interactions. Contextual memory architectures, in contrast, can support long-term personalization, but may be unable to provide real-time knowledge adaptability.The current paper introduces a hybrid model, known as Context-Aware Memory-Augmented Retrieval Architecture(CAMARA), thatcombines dynamic document retrieval with the organized contextual memory that is utility-based and context-aware. Thesuggested system applies adaptive context weighting, memory decay, and relevance-sensitive retrieval totrade-off accuracy in the facts with continuity context-specific to the user. As a result of experimental testing on marketing scenariosof simulated enterprises, CAMARA scores a relevance of 0.93, personalization accuracy of 0.89, as well as minimizes the hallucination rate to 5.8 percent, which is better than standalone RAG and memory-based systems. The framework also hasa lowresponse latency and has a small computational overhead (around 8 percent higher) over baseline RAG. The ablationstudy validates the fact that memory optimization does indeed boost multi-turn coherence at the expense of factual accuracy. These results demonstrate the significance of integrating retrieval and persistent memory into next-generation marketing intelligence systems as an effective and scalable and reliable way to achieve context-aware decision support in the enterprise setting.