A high-level memory API over the document engine: remember, recall, forget and
explain, with temporal validity, supersession, and opt-in consolidation and forgetting.
Learned sparse retrieval: an impact index over learned sparse vectors (SPLADE-style
vocabulary-sized vectors, or any sparse term weights) with exact dot-product top-k.
Overwrites the stack region below the caller’s frame so secrets left there by callees
(Argon2 temporaries, a cipher value before it moved to the heap) do not linger. Best
effort: the language guarantees nothing about frame layout, so the key-zeroization test
in this module measures the result rather than assuming it.