Memory Deletion — References

The obligation that drives erasure, the standards for making data unrecoverable, and the research on why removing content from a derived artefact is harder than removing it from a store.

The obligation

• Regulation (EU) 2016/679 (GDPR) — Article 17, the right to erasure, including the obligation on a controller that has made data public to take reasonable steps to inform others processing it; Article 19 on notifying recipients; and Article 12(3) on the deadline for responding to a request.
• California Consumer Privacy Act, as amended by the CPRA — the deletion right and the obligation to direct service providers and contractors to delete, which is the multi-system propagation problem stated in law.

Making data unrecoverable

• NIST Special Publication 800-88 Revision 1: Guidelines for Media Sanitization — Richard Kissel, Andrew Regenscheid, Matthew Scholl and Kevin Stine, National Institute of Standards and Technology, 2014. The distinction between clearing, purging, and destroying, and the principle that a sanitization claim requires verification rather than assertion. Cited by name; the publication's URL has changed with NIST's catalogue.
• The Log-Structured Merge-Tree (LSM-Tree) — Patrick O'Neil, Edward Cheng, Dieter Gawlick and Elizabeth O'Neil, Acta Informatica, 1996. Why a delete in many modern stores writes a marker and the original bytes survive until compaction reaches them. Cited by name.
• Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs — Yury A. Malkov and Dmitry A. Yashunin, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. The graph structure behind most vector indexes, and why removal from it is a rebuild rather than a deletion. Cited by name.

Why derived artefacts retain what was erased

• Machine Unlearning — Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie and Nicolas Papernot, IEEE Symposium on Security and Privacy, 2021. The cost of removing a specific contribution from something that was derived from many, and the partitioning strategies that make it tractable. Cited by name.
• Towards Making Systems Forget with Machine Unlearning — Yinzhi Cao and Junfeng Yang, IEEE Symposium on Security and Privacy, 2015. The earlier formulation of the same problem in terms of lineage through derived state. Cited by name.
• Extracting Training Data from Large Language Models — Nicholas Carlini, Florian Tramèr, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Úlfar Erlingsson, Alina Oprea and Colin Raffel, USENIX Security, 2021. Why content absorbed into a model's parameters is not reachable by any store-level deletion. Cited by name.

No verified reference is recorded for erasure propagation through agent memory derivation graphs, or for regeneration of consolidated items after a source is erased. Neither has a published treatment that this entry could point to.