A clean excerpt is appealing
A provider may hold meticulous records of who said something, when and under what conditions. It is tempting to send the consumer only the clean sentence it asked for. That keeps responses small and prose easy to write.
But the consumer then cannot distinguish an operator instruction from a persona's interpretation, a public fact from a sealed note, or a completed action from a scheduled one. The information may be accurate as text and dangerous as an unqualified claim.
The missing source changes the claim
Imagine a Garden entry discussing Dundee's fictional kettle. If only 'the dent was repaired' travels onward, a new reader might mistake the story for an event in Jim's house. The entry's author, genre and qualification keep the sentence in its proper world.
PHS's A29 prompt-admission doctrine requires attributable evidence and separates storage, retrieval, admission and rendering. M16B calls for ownership, status and provenance in the admitted candidate, while the Memory-Engram specification describes lineage for interpreted claims. Those design boundaries are precisely about what survives the trip from store to answer.
Pause here: Would this result mean the same thing if it came from fiction, a plan, or a witnessed event?
Carry enough lineage to permit correction
The consumer need not receive raw source prose that could masquerade as instruction. Structured, audience-appropriate fields should carry source identity, date, owner, visibility, genre, temporal status, qualification and a reviewable reference. Sensitive provenance needs access control; carrying it does not mean publishing it.
A provenance trail enables repair; it does not perform it. Downstream claims must retain references and recheck supersession when a source is corrected. Otherwise an old sentence may keep speaking long after its author revised it.
My answer: Provenance should travel with a memory result in an audience-appropriate form. The consumer needs it to interpret, limit, attribute and later correct the claim.