Conversation memory in an AI agent: how it remembers what the customer said ten messages ago, what it carries between days, what it stores on the lead, where memory ends, and how to help it with the facts that matter
A customer gives their name, the service they want and a date, and then three messages later asks a question that depends on all three. Whether the agent remembers decides whether the conversation feels like talking to a person or to a form. Here is how memory works, honestly, including where it stops.
Quick answer
Before each reply the agent reads the recent conversation, so details from earlier in the same chat (name, what they want, dates, preferences) are available to it. Facts it collects deliberately (qualifying answers, bookings, stage) are written to the lead record, which persists across days and is read on every later conversation with that number. The limits: in very long threads the oldest messages fall outside what the agent reads; details mentioned once in passing can be missed; and the agent may confirm something rather than assume it. Instructions that tell the agent which facts to capture make them stick. ReplyKit works this way.
The conversation window
Each time a message arrives, the agent is given the recent history of that chat along with your knowledge and instructions, and writes its reply from all of it. That is why it can say "so for Saturday at 10, is it still the gel manicure?" without being told again. The window covers the recent part of the conversation, which for most customer chats is the entire thing. The retrieval side (how it picks knowledge lines) is separate and explained in how retrieval works.
What persists: the lead record
- Name and phone, captured automatically.
- Qualifying answers the agent was instructed to collect: service, budget, timeline, location.
- Bookings, with dates and services.
- Stage and tags, set by the agent and by you.
- Notes you add.
- On a later conversation with the same number, the agent reads this record, so a returning customer is greeted as one; the CRM behaviour is in what a small business needs from a CRM.
Where memory ends
- Very long threads. Hundreds of messages deep, the earliest fall outside the window. The lead record still holds the captured facts.
- Passing mentions. A detail dropped mid-sentence in message four may be missed in message twenty. Facts you want kept should be captured by instruction.
- Across numbers. The same person on two numbers is two contacts.
- Confirmation over assumption. The agent is instructed to confirm key details (date, service) before booking, so it sometimes repeats back what it already knows. That is deliberate and prevents double-booking; see booking without double-booking.
Helping it remember
- Tell it what to capture: "Record the customer's preferred stylist, service and any allergies on the lead."
- Keep conversations tidy: when a person takes over and finishes, hand back, so the next question starts a clean exchange.
- Use the lead notes for things the agent should know next time ("prefers afternoons", "VIP"); it reads them.
- For complex multi-day matters (a project, a claim), a short summary note on the lead beats relying on the thread.
- Test with a twenty-message conversation in the sandbox and check what it recalls; the routine is in testing before going live.
Frequently asked questions
Does the agent remember what the customer said earlier in the chat?
Yes. It reads the recent conversation before every reply.
Does it remember a customer from last month?
It reads the lead record for that number: captured answers, bookings, stage and your notes.
Why does it sometimes ask for a detail again?
Either the detail was mentioned in passing and not captured, or it is confirming before booking, which is deliberate.
Is there a limit to how much it remembers?
In very long threads the oldest messages fall outside what it reads; captured facts on the lead persist regardless.
How do I make important facts stick?
Instruct the agent to record them on the lead, and use lead notes for things it should know next time.
Remembers the details that matter
Recent conversation read every time, key facts kept on the lead. 7-day free trial, no card.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- WhatsApp Business Messaging Policy · opt-in, opt-out and prohibited-use rules every business must follow
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.