GDPR erasure requests and AI agent memory

A valid erasure request reaches everything an AI agent holds about the person: its memory, conversation history, caches, retrieval indexes and the content in its logs. GDPR's right to erasure has exceptions, including legal obligations and legal claims, which often cover the core audit record. The design answer is to erase personal content while keeping a lean, lawful event trail.

What does the right to erasure require?

GDPR gives people the right to have personal data about them erased without undue delay where one of several grounds applies. In substance, the grounds include:

• the data is no longer necessary for the purposes it was collected or processed for;
• the person withdraws consent and there is no other lawful basis;
• the person objects and there are no overriding legitimate grounds, or they object to direct marketing;
• the data has been processed unlawfully;
• erasure is required by a legal obligation.

The right does not apply where processing is necessary for, among other things, compliance with a legal obligation, the establishment, exercise or defence of legal claims, or certain public interest purposes. Where data has been made public, the controller must take reasonable steps to inform others processing it. And the controller must tell recipients of the data about the erasure, unless impossible or disproportionate.

Who must act?

The controller, within the statutory time limit. Processors, including agent platform and model providers holding the data, must help and must erase on instruction. The process side, including what an agent can do to help, is in how to handle erasure requests with an AI agent, and the narrower question of whether an AI agent can handle a right to erasure request is answered separately.

Where does personal data hide in an agent?

Erasure in a database means deleting rows. An agent spreads personal data across more places, many of which no one designed as storage:

1. Memory. Facts about the person stored for future sessions. Our explainer on what agent memory is describes the forms it takes.
2. Conversation history. Transcripts of interactions.
3. Working caches. Copies of documents and tool results kept to speed up later steps.
4. Retrieval indexes. Searchable representations of documents, which may contain or encode personal data from them.
5. Agent-written notes. Summaries and inferences written back into business systems.
6. Logs. Action records, prompts and responses.
7. The model provider. Prompts and outputs it retained under your contract.
8. Evaluation sets. Real cases kept for testing and certification.

An erasure process that clears the CRM but leaves the agent's memory intact has not honoured the request. The agent may quote the erased data back to the person, or to someone else, next week.

How do you erase without breaking the audit trail?

Three questions decide what goes and what stays.

Is a ground for erasure met? If the person is still a customer and the data is needed to service the account, most data stays. Erasure requests from former customers, or about data processed for optional purposes, are where most of it applies.

Does an exception apply? Financial firms are often required by law to keep records of transactions, communications and decisions for set periods. Where agent action records are kept to meet such obligations, or for potential legal claims, the exception may cover them. Document which records fall under which exception.

Can the purpose of retention be met with less? Usually yes. The audit trail needs to show that an action happened, which agent did it, which rule decided and who approved. It rarely needs the customer's message text or a copy of their bank statement. Erase the content and keep the pseudonymised event.

The last row is specific to agents. Without a flag, an agent that meets the person again, or reads an old document about them, may quietly recreate what was erased.

Tiered logging makes all of this practical. Our piece on GDPR storage limitation and AI agent memory explains the retention design, and the access side is covered in GDPR data subject access requests and AI agent logs.

What will a supervisor or auditor ask?

• When a person asks for erasure, which agent stores do you clear?
• Show us an erasure request completed across memory, caches and logs.
• What do you keep, and under which exception?
• How do you stop agents recreating erased data?
• Did you tell your model provider?

What does the right to erasure not require?

It does not require erasing data you must keep by law or for legal claims. It does not apply where no ground is met, such as data needed to perform a live contract. It does not require retraining a model you never trained on the data; qualification-based agents that do not change model weights avoid that problem entirely. For the full map, see AI agent regulation, mapped to controls.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.