Teach it once.Ask it anytime.

Duplex looks up everything your team has taught it.

Learned answers come back with their source, the date they were taught, their version and when they expire, next to related questions, documents, facts and standing rules. It reads your own memory and your organization's, side by side.

A calibrated gate chooses one of four paths.

Duplex commits only when its confidence clears a threshold tuned to the risk you set, and can be held back until it has calibrated on real turns. Live questions go to a tool, the rest to your model, and anything a rule forbids is refused, with the rule named.

Served from memory, or by the model in your client.

A committed answer arrives instantly, with no model call and a source line your people can check. A deferred turn is answered by the model your AI client already uses: Duplex calls no model and holds no model key.

Every approved reply becomes memory.

Model answers stay unverified until a person approves them. A correction creates a new version instead of overwriting the old one, a rule takes effect only after the user confirms it, and forgetting is complete, with an erasure certificate to prove it.

Your organization stays in control.

Everyone reads the organization's shared memory; only its admins write it. Every decision lands on a signed audit ledger, and usage is counted in governed decisions, never in tokens.