Key answer
If old goals/material remain in context and the new request leaves objects or scope unclear, AI may mix them. Signs include earlier project numbers, client names or styles, or rejected proposals retained as premises. This can happen before a conversation becomes long.
Full explanation
Shi et al. added irrelevant, answer-preserving information to math word problems and found interference. GSM-IC study Liu et al. found position effects in multi-document QA and key-value retrieval. Lost in the Middle Neither measures a universal number of turns before topic contamination.
Start with minimal clarification: current subject, current sources and still-valid old requirements. History helps wording changes to one report; a new client, dataset or value judgment needs explicit boundaries.
Next, write Project B's manual from this attachment and requirements. Project A's numbers, people and conclusions do not apply. Retain only the beginner audience and English-language preference. Complete the manual directly and mark missing facts.
Spot-check likely contamination: entities, dates, units, terms and goals. If old assumptions recur, consolidate current requirements and start a new conversation. A scope statement guides information use but does not delete history; storage, memory and privacy need separate product controls.