The word “if” matters. It changes the claim from an assertion about the world into a hypothetical premise for this conversation. Similar traps are “I might,” “my colleague lives in,” quoted text, and a tool result that merely suggests an address. I would first recover the exact message span that caused the memory write and its surrounding turns. Was this an explicit user assertion, a question, a conditional, a quotation, or an inference made by the assistant? A confidence number alone will not repair a missing speech act.

I would keep conversation evidence separate from durable personal facts. An extraction candidate should carry the subject, predicate, value, original quote and turn ID, speaker, modality or conditional scope, event time, and whether the user explicitly asserted it. The promotion policy can accept “I moved to London last week” as a candidate for durable residence, while keeping “if I move” in episodic conversation history as a hypothetical. For consequential facts, ask the user or rely on an authoritative profile field before acting. Even a direct assertion can become stale, so provenance and a way to correct it still matter.

If someone says, “Just use an LLM to classify hypotheticals,” I would use that classifier as a gate, then measure its false promotions with adversarial pairs that differ by a few words. Include counterfactual planning, other people's biographies, negation, sarcasm and corrections. Inspect the write decision and the later action, not only extraction accuracy. A small false positive rate can compound when a long-lived fact is injected into many future prompts. The safer default for uncertain candidates is no durable write, with a way for the user to confirm a useful memory.

At read time, do not quietly turn a conversation snippet into a personal profile. Show the provenance to the assistant and give a conditional statement conditional status. If wrong facts have already propagated into summaries, invalidate descendants and correct the user-visible profile. Research on provenance-aware agent memory motivates keeping events and curated facts distinguishable. The classification and promotion rules here are a design proposal, not a rule established by that paper. The memory expired yesterday. Why does the agent still act on it today? is about a valid temporary fact outliving its expiry. Here the fact was never asserted in the first place.