Structural Forgetting in AI Memory Systems: A Two-Phase Framework Unifying Index Blocking, Output Gating, Selective Power Gating, and Key Revocation

This paper redefines “forgetting” in AI systems beyond physical deletion, proposing structural forgetting as an operational design space for reducing the reachability and reconstructability of target information across system layers. We present a unified framework with four core mechanisms: (1) index blocking (route/index invalidation), (2) output gating (generation suppression), (3) selective power gating (sleeping caches/KV and related memory surfaces), and (4) key revocation (crypto-erasure via irreversible key destruction). We further introduce a two-phase trace model—a reversible phase for default mitigation and an irreversible phase for escalation—enabling practical trade-offs among safety, legal compliance (e.g., deletion requests), and energy efficiency. To make these designs testable, we propose an evaluation protocol combining ML-unlearning-style leakage tests, output-gating A/B assessments, power-gating simulations, and key-revocation drills, alongside a composite metric (Forgetting Efficacy, FE) that integrates leakage suppression, reversibility, energy reduction, and operational cost. The proposed framework supports realistic “forget-without-deleting” operation while reserving irreversible actions for high-stakes scenarios, providing implementable guidance for safe and accountable AI deployment. This is a preprint. The paper has not undergone peer review.