Semantic Resurrection Bounds Verifiable Deletion: A Measured Characterization of the Identity Ceiling in Agent-Memory Forgetting

Abstract

Cryptographically verifiable deletion in agent memory is an active area: dependency-aware unlearning, cascade repair, and crypto-shredding with deletion receipts are all established. This note addresses a narrower question that survives that prior art: what bounds verifiable deletion when identity is semantic rather than syntactic? A digest-level non-membership proof is airtight over bytes, but a paraphrase of a forgotten fact carries a different digest and the same meaning. We measure that seam on 703 real paraphrase pairs and report a characterization rather than a system: semantic resurrection detection separates paraphrases from distinct facts at AUROC 0.951, but the operating point is governed by a recall-precision frontier that is the same false-merge/false-split wall documented for belief identity, now in deletion dress. We compose a digest-level accumulator proof with a semantic centroid guard (100% non-membership verification retained; 27.6% resurrection catch at the chosen operating point, ~67% at a looser one) and extend a declared-edge cascade to implicit descendants (87.3-100%) at a measured precision cost (16.0-38.7% false cascade), while opaque reasoning-only descendants remain at 0% under every method tested. We argue the residual is an operating-point choice, not an irreducible bound, and that the irreducible part (~5%) is set by the identity ceiling itself.