Selective Memory Deletion and User-Controlled Memory Erasure in Generative Artificial Intelligence: A Lifecycle Governance Architecture for User-Directed Forgetting

The increasing adoption of persistent memory in generative artificial intelligence (AI) systems has enabled personalized interactions, cross-conversation continuity, and long-term user–AI collaboration. However, the development of conversational memory has created an equally important but comparatively underdeveloped design question: how can users selectively remove information that they no longer want an AI system to remember, retrieve, or use? Users may wish to delete memories because they are emotionally uncomfortable, personally sensitive, inaccurate, outdated, potentially dangerous, or likely to create unwanted consequences. They may also wish to remove information for no reason other than personal preference. Existing memory-management mechanisms provide partial deletion capabilities, while research on machine unlearning investigates the removal of information from model parameters. Nevertheless, user-directed deletion of conversational memory raises a broader architectural problem involving the identification of target memories, the tracing of derived representations, the propagation of deletion requests across memory components, and the verification that deleted information is no longer available through future retrieval. This paper proposes Selective Memory Deletion and User-Controlled Memory Erasure (SMD-UCME), a conceptual architecture for governing the selective removal of user-specific information from generative AI memory systems. The framework introduces five complementary deletion mechanisms: individual memory deletion, category-based deletion, semantic memory deletion, risk-oriented memory identification, and emotion-oriented memory identification. Its central contribution is the integration of user-defined deletion intent with memory provenance tracking, dependency-aware deletion propagation, retrieval exclusion, and post-deletion verification. The architecture distinguishes the deletion of an AI memory representation from the deletion of the original conversational record, and further distinguishes both operations from model-level machine unlearning. A formal memory-dependency model, deletion-policy representation, operational workflow, system invariants, and experimentally testable hypotheses are developed. The framework also addresses consent, false identification, partial deletion, distributed storage, cached representations, external sources, and the limits of deletion guarantees. Building on the author's previously published Memory-Isolated, History-Preserved, and Future-Integrable Conversations (MIHFI) architecture, this study extends user control from the conditions under which memories are retrieved and integrated to the conditions under which previously retained memories can be selectively withdrawn. The paper is conceptual and does not claim an implemented system or experimentally demonstrated performance. It establishes a testable research direction for making user-directed forgetting an explicit component of conversational memory lifecycle governance.