WhatsApp AI agent memory: what should it remember?

Conversation state records the task in progress: a selected service, an unanswered question or a proposed date. It should change when the customer corrects the request.

Preferences may outlast a conversation, but should have a clear purpose and a way to update them. A one-off request for an English reply does not establish a permanent language preference.

Business facts belong in their authoritative system. An agent should refresh a price, booking status or order state before acting. A remembered answer is not evidence that the fact is still current.

For each retained item, record value; source; time observed; scope; expiry or refresh rule; correction owner. This is a design worksheet, not a description of a built-in Waaru memory database.

Fictional booking record: preferred branch = city; source = customer reply; scope = current request; selected time = unconfirmed. When the customer says "Actually, use the airport branch," replace the preference and recheck availability. Do not carry the old slot into the new location.

"The customer says they paid" and "the payment system confirms payment" are different records. Summaries should preserve that distinction. Otherwise a later action can treat an unverified statement as permission to fulfil an order.

Store the unresolved question explicitly. A summary such as "payment sorted" is too vague for an action that depends on settlement. The tools and actions guide explains why permissions must remain outside the model's judgement.

When a customer corrects a date or preference, the next reply should use that correction. An old summary should not make them explain it twice. Make the change visible to whoever takes over, and define which records staff can amend and how those changes are recorded.

Choose retention according to the actual purpose and applicable policy. Do not keep sensitive details merely because they could be useful later. A request to delete data needs the organisation's verified deletion process; a conversational promise does not prove stored copies were removed.

Use fictional records to test a changed date, a repeated customer after a long gap, two simultaneous tasks and a handoff to staff. Confirm that one customer's context never appears in another conversation.

Also test stale business data: a price changes after the first answer, or an appointment is cancelled elsewhere. The next response should use the updated source. Read how to test an AI agent for the wider test process. Treat long-term memory or automatic learning as unverified until the deployed implementation proves it.