Retell Memory retains context across repeat contacts. We examine how to test a new-order scenario when a customer changes delivery instructions.
On October 8, 2026, Retell AI introduced Memory for voice agents. It maintains a brief for each contact, with separate settings for saving and using that context. Completed phone calls and SMS chats update the brief when saving is enabled. Web calls and web chats do not update it automatically.
An online store could begin by testing a repeat inquiry. Suppose a customer asked for their first order to be delivered to the office yesterday and calls about a second order today. In this scenario, the stored preference should prompt an address check, rather than being carried over to another purchase without confirmation.
The team needs to distinguish order-specific instructions from a customer's general preferences. A request for an evening callback, for example, may apply only once. The acceptance criterion is concrete: the agent confirms the current preference for the correct order, and the integration records the agreed response.
Create a test customer with two different orders. Change the delivery instructions in the first conversation, then provide a different address in the second and check which order receives the change. Compare what the customer said, the conversation brief, and the record in the store's test system. This can reveal an old preference being carried over to a new order.
Define the source of current status separately: the workflow should query the order system during the inquiry. Before confirmation is complete, irreversible actions should not rely solely on a previous conversation. Assess the pilot by whether it selects the right order and records the correct changes, separately from how smoothly the agent speaks.
In the proposed workflow, the team should require confirmation of the current order and new instructions before recording changes.
Check whether it selects the correct order and whether the recorded change matches the customer's latest confirmation. Evaluate conversational fluency separately.
Voice agents, telephony and integration with your business systems — using your own infrastructure or international platforms.
Explore our services ↗The Vapi platform now allows discarding low-confidence Soniox transcriptions and controlling latency in voice agent simulations.
Read articleDeepgram introduced pause and pronunciation controls directly in the text for Flux TTS. The update allows inserting silence and specifying IPA transcriptions.
Read articleThe new models support emotion control via tags and over 90 languages. According to ElevenLabs, the median time to first speech is around 150ms.
Read article
