How would you evaluate a personal AI for invented memories and revoked consent?

A practitioner proposal argues that personal AI memory systems should be scored on retrieval and final-answer stages separately, since an invented memory can originate in either, and should be evaluated on abstention alongside accuracy so a system that refuses everything does not appear perfect. The proposal cites LongMemEval (arXiv 2410.10813), which tests information extraction, multi-session reasoning, temporal reasoning, knowledge updates and abstention and splits memory into indexing, retrieval and reading, and TOFU (arXiv 2401.06121), which uses 200 diverse synthetic author profiles with a forget set, as the nearest setup for testing revoked consent by treating a revoked fact as forget-set data and checking it never returns in answers or the store.

Two baselines that cover parts of this: LongMemEval \ 2410.10813\ LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory https://arxiv.org/abs/2410.10813 tests “information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention”, and splits memory into “indexing, retrieval, and reading”. That split answers whether to score retrieval and the final answer separately: I think you should, since an invented memory can come from either stage. For revoked consent, TOFU \ 2401.06121\ TOFU: A Task of Fictitious Unlearning for LLMs https://arxiv.org/abs/2401.06121 is the nearest setup: “200 diverse synthetic author profiles” with a “forget set that serves as the target for unlearning”. You can treat a revoked fact as forget-set data and check that it never comes back in answers; not only that, but it is gone from the store. One suggestion from my side: score abstention on questions the system should not know, together with accuracy on ones it should. Otherwise, a system that refuses everything looks perfect on invented memories.