When the Assistant finds two answers
Connect shows you both sides with their origin and their age, and does not pick a winner on your behalf. Where precedence is designed — memory tiers, autonomy scopes — the narrower one wins automatically and that is not a conflict. A real disagreement between two statements of fact is a business decision, and a business decision belongs to a person.
Designed precedence is not conflict#
Two of Connect's hierarchies resolve by rule, deliberately, and neither should be mistaken for the Assistant choosing sides.
| Hierarchy | Order | What it means |
|---|---|---|
| Memory | workspace → channel → endpoint → contact | A note about one contact overrides the workspace-wide one for that contact only |
| Autonomy | contact → endpoint → channel → workspace, narrowest first | A rule set against one contact beats the channel's |
In both, the narrower statement is more specific rather than more recent, and that is the whole intent: you set a general policy once and carve out the exceptions that matter. An answer that reflects a contact-tier memory over a workspace one is following your instruction, not overruling it.
The conflicts that are real#
- Knowledge against knowledge
- Two sources a workspace supplied that state different things — commonly an old document nobody retired and a newer one. Both are grounded, both are yours, and neither is automatically right.
- Memory against knowledge
- Somebody told Connect one thing; a document says another. The memory is usually newer and narrower; the document is usually more authoritative. That tension is exactly why nothing is resolved silently.
- A conversation against the file
- A customer says an arrangement is one way and your own recorded position says otherwise. This one is not a data problem at all — it is a commercial disagreement wearing a data problem's clothes.
- A fact against its own source
- An extracted fact that no longer matches the source it was drawn from, because the source changed. The fact is the thing to re-check, not the answer built on it.
What Connect does about it#
It surfaces the disagreement rather than averaging it. An answer over conflicting material names both statements, says where each came from, and says which is more recent — and then stops, because choosing between two things your own business has said is not a decision an assistant is entitled to make quietly.
It also does not act on the ambiguity. A proposed action that depends on which side is right arrives with that dependency stated, and the confirmation is where you resolve it. If the question is important enough to hold up work, it goes to the queue a person works rather than sitting in a chat window nobody reopens.
Settling one properly#
Decide which statement is right — that part is yours and cannot be delegated.
Result You now have a resolution rather than two candidates.
Fix it where it lives. A stale knowledge source is retired or replaced; a wrong memory is forgotten at the tier that holds it; an outdated fact is corrected.
Result The next answer is drawn from corrected material, so nobody has to remember the resolution.
Write the decision down if the reasoning matters, as a memory at the tier it applies to or as a standing instruction if it is a rule rather than a fact.
Result Behaviour changes, not just the answer. Making a correction actually change behaviour is the long version.
Correcting the answer in conversation and stopping there is the failure mode to avoid. The engine that drafts replies does not read your chat, so a resolution that lives only in a tab will be contradicted by the next draft it writes.
Questions#
Does the newest information always win?
No, and it should not. Memory resolves by specificity rather than date, and a conflict between two sources is surfaced with both dates shown so you can weigh them. Automatic recency would quietly promote every hasty note above every considered position.
What if the customer is the one contradicting our records?
That is a commercial matter, not a data-quality one. Connect will show you what it holds and where it came from; what your business does about the difference is a decision for a person, and recording the outcome is what stops it recurring.
Can I make one source authoritative for everything?
There is no single 'trust this above all' switch documented, and this page does not invent one. In practice the same effect comes from retiring what is wrong, so there is nothing to conflict with, plus a standing instruction where a rule genuinely needs stating.