DEV Community

RingBot AI
RingBot AI

Posted on Edited on

What an Agent Should Remember — and What It Must Forget

Memory is the most requested capability in agent deployments and the easiest to get ethically wrong.

The pitch is appealing: the agent remembers that Mrs Chen prefers Thursday mornings, so the next call starts warm instead of cold. The failure is equally real: the agent remembers something it should not have, tells the wrong person, and your business now owns a data incident.

What memory is actually for

Before deciding what to remember, decide what problem memory solves. In practice there are three:

Continuity. The customer should not repeat their order number, their address, or their issue on every call.

Preference. The same customer prefers the same slot, the same language, the same channel.

Context. Something happened last time — a complaint, a refund, a promise — and the next conversation should not pretend otherwise.

If a fact does not serve one of those three, it probably does not belong in memory.

What is worth remembering

  • Verified contact details — the number and email the customer confirmed
  • Stated preferences — preferred time, language, channel, contact method
  • Open commitments — a callback promised, a replacement arranged, a refund in progress
  • History that changes the answer — a previous complaint, an exception already granted

Each of these makes the next interaction shorter or safer. That is the test: does remembering this reduce work or reduce risk?

What must never be inferred

This is where most systems go wrong, and it is not a technology limit — it is a decision.

  • Anything about health. Even if the customer mentions it on a call, it should not become a stored attribute.
  • Anything about finances beyond what is needed to complete a transaction.
  • Anything about family or household inferred from context.
  • Anything about emotional state or personality. "Frustrated customer" is not a fact about a person; it is a judgement about a moment, and storing it biases every future interaction.
  • Anything a third party said about the customer.

The rule of thumb: store what the customer would recognise and agree with if you read it back to them.

If reading the memory aloud would make them uncomfortable, it does not belong there.

The consent question

Memory has a distribution problem: the customer gives information in one conversation, and it is used in another. That is a different act from keeping notes for the current call.

Three positions:

  • Default off. Memory only with explicit opt-in. Safest, and the reason most vendors do not offer it as a default.
  • Default on with notice. Acceptable where the notice is genuinely clear, not buried.
  • Default on silently. Indefensible, and increasingly illegal.

Whichever you choose, the customer must be able to see what is stored and have it removed. That is not a nice-to-have; in several jurisdictions it is the law, and in all of them it is the thing people ask about once they realise memory exists.

The four controls to be able to demonstrate

If someone asks how memory works in your deployment, you should be able to show:

1. What is stored, per customer, in plain language. 2. How it got there — which conversation, and whether the customer said it or the system inferred it. 3. Who can see it — and specifically whether frontline staff can read the whole record or only what they need. 4. How it is deleted — and whether deletion removes the derived copies too, not just the visible record.

The fourth is the one that fails audits. "We deleted the profile" is not the same as "we removed every derived artefact".

The staleness problem

Memory that is never reviewed becomes wrong.

A preference recorded eighteen months ago may be a preference no longer held. An open commitment from a resolved case reads as unfinished business. A stored address that has changed produces a delivery to the wrong place.

Two disciplines prevent this:

  • Age matters. Recent information can be used freely; older information should be confirmed before it is acted on. "Is this still the right number?" costs four seconds and prevents a category of error.
  • Contradictions need resolution. If the customer says one thing and memory says another, the new statement wins — and the old one should be superseded, not left alongside.

The honest case against memory

For many small businesses, memory is not worth the risk.

If your calls are mostly new enquiries — quotes, bookings, hours — the value of remembering is low and the liability is real. A customer calls once, books, and never speaks to you again until the appointment. Continuity adds nothing.

Memory pays off where relationships are long and repeated: account management, ongoing service, recurring appointments. In those cases it is genuinely transformative. In a one-off transaction business, it is a data protection obligation with no upside.

Being honest about which one you are is worth more than the feature.


Disclosure: I work on RingBot, an Agent-as-a-Service platform where customer memory can be enabled per deployment, reviewed and corrected by the account owner, and deleted on request — including derived data. The split above between what to keep and what to never infer is a policy decision that no platform can make for you.

Related: identity checks, collecting a form by voice, the handoff brief, when to hand off.

Top comments (0)