AI agents with memory: what to remember about a customer

A customer messages on a Tuesday: "hi, it's Marta, from last week." If your agent replies "how can I help you?" as though it's never heard of her, it feels cold. If it replies "Hi Marta! How's the knee after Thursday's session?", that can land beautifully… or badly, depending on who happens to be reading Marta's phone at that moment. That's what memory is: the gap between service that feels personal and service that feels a little creepy.

This piece is for small business owners who have (or are about to have) an agent handling customers on WhatsApp, the website or email. By the end you'll know what an agent's "memory" actually is, what's worth remembering about a customer from one conversation to the next, what it should forget on purpose, and how to decide all of that in an afternoon.

What agent memory is (and isn't)

General-purpose AI assistants now ship with memory built in: they remember you like short answers, or which project you're working on. So the question arrives on its own in any business: "can my agent remember my customers?" It can, but it's worth understanding how, because everything else follows from it.

An agent doesn't "remember" the way a person does. The model itself doesn't change from talking to Marta; tomorrow it's exactly the same model as today. What exists is a note the system saves at the end of a conversation and reads back the next time that person gets in touch. A customer record, essentially. Sometimes written by the agent, sometimes by your team, sometimes by both.

That has three very practical consequences:

• Memory can be read. If you can't open a customer's record and see what the agent "remembers" about them, you have a control problem, not a technology problem.
• Memory can be wrong. If the agent noted something incorrectly on Monday, it will repeat it with total confidence on Friday.
• Memory can be deleted. And sometimes it should be.

Don't confuse this with rereading the whole conversation history every time. We're talking about something smaller and deliberate: a few lines per customer that survive from one conversation to the next.

What it should remember

The test is simple: does this save the customer repeating themselves, or save you a mistake? If yes, it goes in.

• How they like to be contacted. "WhatsApp please, no calls." "Always writes in the afternoon." Ignoring this is what annoys people most, and it's the easiest thing to remember.
• Where you left things. "Quote sent on the 12th, waiting for them to confirm measurements." This is the note that kills "sorry, remind me what we discussed?"
• Decisions they've already made. "Invoice to the company, not personally." "Asked not to receive promotions." That last one especially: sending offers to someone who said no is one of the worst things an automated system can do.
• Service context you'd jot down on paper anyway. The boiler model, their usual size, the fact that the dog is called Coco and gets nervous at the groomer's. What a good long-serving shop assistant would know without looking.

Notice that every item is useful for serving them better, not for knowing more about them. That's the line.

What it should forget (or never write down)

This is where mistakes happen, almost always with good intentions.

Health details, unless health is your business. A physio clinic needs to know Marta is coming in for her knee; a sports shop doesn't. And even for the clinic, that belongs in its clinical management software with proper permissions, not in the WhatsApp agent's memory. GDPR treats health as a special category of data (Article 9) and demands much more care.

Things mentioned in passing. "Sorry I'm late getting back, I'm going through a divorce." A person hears that and politely lets it go. An agent with automatic memory might write it down and bring it up three months later. No customer note should contain anything you'd be embarrassed for the customer to read.

Payment details and documents. Card numbers, bank details, ID numbers: the agent can help handle them in the moment, but there's no reason for it to remember them. They belong in your invoicing system.

Anything that expires. "On holiday until the 15th." "Looking for a two-bedroom flat." True today, false in two months. Store it without a date and one day the agent will ask someone about a flat they bought ages ago.

Rule of thumb: if you wouldn't write it on an old-fashioned paper index card, the agent shouldn't write it down either.

Three boxes: permanent, expiring, never

To decide in an afternoon, sort every type of information into one of three boxes:

1. Permanent (until the customer changes it). Name, preferred channel, billing details, decisions like "no promotions". Reviewed when the customer asks.
2. Expiring. Status of a quote, an open issue, a temporary preference. Give it a date: after 30, 60 or 90 days it's deleted automatically or flagged for review. GDPR calls this storage limitation; it's common sense before it's an obligation.
3. Never. Health (unless it's your line of work), personal matters, payment data, opinions about the customer. That last one matters: "difficult customer" or "always haggles" has no place in a memory that might one day leak into a reply.

Write it on one page. That page is the instruction you give the agent about what to note, and it's also what you show a customer who asks what you hold about them. They're entitled to ask.

When it's better to remember nothing

Not every business benefits from memory across conversations. Think twice if:

• Your customers share devices or numbers. A company line used by three people, the family phone the whole household books from. The agent can't tell who's writing and will greet the wrong person with someone else's details.
• Most of your contacts are one-offs. A question about opening hours, a single order. Keeping notes on thousands of people who won't come back is cost and risk with no upside.
• Your sector is sensitive by nature. Mental health, delicate legal matters, services the customer would rather nobody at home knew about. Here, an agent with no memory that hands over to a person quickly is usually the most respectful option.
• Nobody on your team will review the notes. Unsupervised memory piles up errors over time. If you won't look at it even once a quarter, it's better not to have it.

In these cases a good agent without memory is perfectly fine: it answers well, captures what matters from that conversation and passes it to you. The remembering is done by you, or by your management software.

Where to start

If you already have an agent with memory, open ten customer records at random and read them as if you were the customer. Anything that would make you uncomfortable? Anything out of date? Ten minutes will tell you whether memory is working for you or quietly piling up problems.

If you don't have one yet, start without automatic memory: let the agent read the record you already keep on each customer, but not write to it until your three boxes are clear. You approve what gets stored; the agent uses it.

If you'd like to see how this fits into an agent that handles WhatsApp, web and email from one place, have a look at the receptionist. You decide the three boxes before it goes live.