Skip to content
be-human.ai
← Back to the journal

Context & memory

Customer Preference Notes for AI Agents: What to Keep and What to Leave Out

Learn which customer preferences are worth storing in AI agent memory, such as channel and delivery windows, and which sensitive details to keep out.

Be Human typographic cover with the title: Customer Preference Notes for AI Agents: What to Keep and What to Leave Out
Conceptual illustration created with AI · Be Human

Customer preference notes are short, verified statements about how a customer wants to be served, such as a preferred contact channel, a delivery window or a standing product option. An agent should store one only when it changes a concrete next action and when a new team member would otherwise have to ask again. Anything that reveals health, payment credentials, family circumstances, beliefs or legal exposure stays out of general agent memory, even when the customer shared it in good faith.

Key takeaways

  • Store a preference only when it changes a concrete next action, such as the channel used for an update or the size quoted.
  • Record the source, the confirmation point and an owner for every note so it can be checked and corrected.
  • Keep sensitive categories out of general agent memory, including health details, payment data, identity numbers and inferred personal traits.
  • Separate what the customer stated from what staff assumed, and never save an assumption as a fact.
  • Give each note a review point, because preferences change and a stale note can mislead the agent as much as a missing one.

What counts as a useful customer preference

Useful preferences are operational. They describe how the business should act for this customer, not who the customer is as a person. In most service businesses the list is shorter than teams expect:

  • Preferred channel and hours for updates, such as text messages in the afternoon
  • Language and form of address the customer wants used in writing
  • Standing choices, such as a usual size, finish or service package
  • Site or delivery constraints, such as a building that needs advance notice before arrival (store the need, not the access code)
  • Approval rules the customer has set, such as asking before any change to a quote

A practical test helps here. Ask whether a new employee who had never met this customer would need the fact to serve them well without asking again. If the answer is yes, the fact is a candidate for memory. If the fact only makes a conversation more pleasant, it can stay in the human notes, where a person can judge how to use it.

This distinction matters because an agent does not read a preference the way a person does. A person remembers a preference, weighs it and decides whether to mention it. An agent retrieves what is stored and may place it in a reply, a summary or a draft. A preference that is useful in a human notebook can become an awkward sentence in an automated message if it was never meant to be spoken aloud.

Which details should never enter agent memory

Agent memory is retrieved into later conversations, drafts and summaries, so a detail saved once can appear in places the original record never reached. The OWASP Top 10 for LLM Applications lists sensitive information disclosure as a risk category for this reason, and the NIST AI Risk Management Framework asks organizations to map and manage risks like this before a system is deployed. Keep these out of general agent memory:

  • Health conditions, medications and diagnoses (an operational need such as step-free access can be stored without the reason)
  • Payment card numbers, bank details, passwords, one-time codes and access credentials
  • Government identification numbers, passport details or driver's license details
  • Family situations, separations, custody matters and active disputes or legal exposure
  • Religion, political views, sexual orientation or other personal characteristics, including ones inferred from a purchase or a conversation
  • Staff opinions about the customer, such as labels like difficult or suspicious
  • Location history or home details beyond the delivery or service address

Data protection laws treat several of these categories with extra care. Brazil's LGPD and the EU's GDPR both give sensitive personal data stronger conditions than ordinary contact details, so the legal review belongs in the design of the memory, not in a later cleanup. For Brazilian operations, the ANPD publishes guidance that is worth reading with a qualified advisor before any personal data is written into an automated system. The safest default is simple: if a detail would make the customer uncomfortable to see repeated in an automated reply, it should not be stored.

A decision guide for each candidate note

Run every candidate fact through the same questions before the agent is allowed to write it. Consistency matters more than sophistication here, because a team that applies the same rules every time can review its own decisions later.

  • Did the customer state it directly, or in a message they wrote? If not, record it as an open question for staff, not as a preference.
  • Does the fact change an action the agent can take, such as a channel, a time slot or a quote format? If not, leave it in the human record.
  • Does it fall into a sensitive category? If so, store only the operational need, or keep it out entirely.
  • Would the customer be surprised to see the agent use it in a reply? If yes, ask for confirmation before saving.
  • Who owns the note, and when will someone review it? If nobody can answer, do not save it yet.

A note that fails any of these questions is not useless. It is simply not ready for agent memory. Keeping it in a staff record, or leaving it unrecorded, is often the right outcome.

Writing a preference note that stays useful

A good note is short, specific and traceable. Each one should carry the statement itself, its scope, its source, the point at which it was confirmed, an owner and a review point. When a note lacks any of these, the agent can still use it, but nobody can tell later whether it deserves trust. A practical sequence for a first set looks like this:

  • List the five or six preferences staff repeat most often in the first weeks of a customer relationship.
  • Keep only the ones that pass the decision guide, and rewrite each as one plain sentence.
  • Attach the source and the person who confirmed it, so a later dispute has a trail to follow.
  • Set the scope: this customer only, this account, or a shared team rule. Shared rules need their own owner.
  • Agree on a review point, such as when the account is reviewed or when the customer's contact details change.
  • Test the agent with three or four realistic requests and confirm it uses the note correctly and nothing else.

Illustrative scenario: a furniture repair studio hears at the front desk that a customer prefers text messages to phone calls and wants pickups after four in the afternoon. The stored note reads: prefers text updates; pickups after 16:00 when possible; source, front desk conversation; owner, studio manager; review at each new order. During the same call she mentions a family illness that is delaying her decision. That detail is not saved. If the studio needs to pause the order, staff record only the pause and the agreed new date in the job file, where it belongs.

Who can write, change and delete preference notes

The agent should be able to propose a preference, not confirm it silently. Writing rights belong to a short list of roles, usually whoever manages accounts, and the agent's proposals wait in a queue until someone approves them. This keeps a single misheard sentence from becoming a permanent rule. Customer facts and staff opinions also need separate fields, so that a judgment about a person never looks like something the person told you. The approach is set out in more detail in our article on customer facts versus staff opinions.

Corrections should follow the same path. When a customer changes a preference, the correction should replace the old note at its source, not sit beside it as a contradicting line. When a customer asks for removal, the team needs a documented deletion path that also covers copies the agent may have placed in summaries. Our guide on deleting agent memory separates correcting, expiring and removing as distinct actions, which helps when a request is not clear at first.

How Be Human handles preference memory in a managed setup

Be Human starts each engagement by defining one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion, so preference notes are scoped from the first day rather than added later. In managed operations, the team reviews which notes the agent writes, checks permissions when staff roles change, and treats a wrong or sensitive note as an incident that is corrected at the source record. No memory design can be guaranteed in advance, so the work is judged against the criterion agreed for that specific role.

Reviewing preference notes over time

Preferences go stale faster than most business records, because people change channels, addresses and tastes without telling anyone. Set a recurring review that checks whether each note is still true, whether two notes now conflict, and whether any note has drifted into a sensitive category. Sample a few agent replies each month and look for preferences applied where they no longer fit. Stale context is a frequent cause of misleading replies, and our article on stale context in AI agents explains how to catch it early. A recurring review also shows which notes earn their place and which should be retired.

If you are not sure which customer preferences your team should hand to an agent first, the free Be Human AI diagnostic is a sensible place to find the first task worth handing over.

Frequently asked questions

What is customer preference memory in an AI agent? Customer preference memory is a small, reviewed set of facts about how a customer wants to be served, such as channel, timing, language and standing choices. It differs from chat history because each note is deliberately written, scoped and owned. The agent uses the note to act correctly, not to reconstruct past conversations.

Should an agent store everything a customer says? No. Most of what a customer says in a conversation is context for that exchange and does not need to persist. Store only facts that change a future action, that the customer stated themselves, and that pass the sensitivity check. The rest can stay in the human record or be discarded.

How do I handle a preference that a customer later changes? Replace the old note at its source record rather than adding a contradicting line beside it. Record the new source and confirmation point, and remove the old wording from any summaries the agent produced. If the change affects other notes, review those in the same pass so conflicts do not linger.

Can an agent ask a customer for permission to save a preference? Yes, and for anything the customer might not expect the business to remember, it should. A short confirmation, such as asking whether text updates are fine from now on, is usually enough. The answer, with its source and the time it was given, becomes the note. That step also tells the customer clearly what the business keeps.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →