Does GDPR apply to what an AI agent remembers?

Yes, whenever the memory contains personal data, and it usually does. Transcripts, summaries of past conversations, notes the agent writes for itself and embeddings built from customer records are all processing of personal data. GDPR's principles, the person's rights and your record-keeping duties apply to each one, just as they do to your CRM.

Which kinds of agent memory count?

An agent can remember in several places, and each is a separate store of personal data:

• Conversation history kept so the agent can pick up where it left off.
• Long-term memory: facts the agent saves about a customer, such as "prefers email" or "disputed a charge in spring".
• Summaries the agent writes of cases or calls.
• Embeddings of documents and messages, stored for search. Converting text into numbers does not take it out of GDPR if it can still be linked to a person.
• Logs and traces of what the agent read and did.

GDPR applies to all of them if they relate to an identified or identifiable person. The organisation that decides why the agent remembers is the controller. A model or platform provider that stores memory on your behalf is usually a processor. The design choices behind this are argued in what an agent remembers is a data decision.

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

What does GDPR require of an agent's memory?

The Article 5 principles do most of the work:

• Purpose limitation. Memory kept to resolve a complaint should not quietly become memory used to sell to the same customer.
• Data minimisation. Remember what the job needs. An agent that saves every detail "in case it helps" breaches this by design.
• Accuracy. An agent's notes are inferences. If it records "customer is in financial difficulty" on thin evidence, that inaccurate fact will shape later actions.
• Storage limitation. Each memory store needs a retention period and deletion that works.

Then come the rights. A person who asks for access must receive what the agent remembers about them; a person who asks for erasure must have it removed from every store, including embeddings and summaries. Memory that builds a picture of someone's preferences or behaviour can also be profiling, covered in GDPR profiling and AI agents.

What may the agent remember, and what stays with a person?

An illustrative example

Take a utility company's support agent with long-term memory switched on. Over months, it builds notes such as "struggles to pay", "lost a parent", "speaks to us angrily". A customer submits an access request and receives none of this, because the access process only searches the customer database.

The company has two problems: it holds sensitive inferences it never decided to collect, and its rights process misses a whole store. It responds by limiting what the agent may write to memory to a short list of categories, marking each item with its source, setting a 90-day expiry on case notes, banning health and bereavement notes from memory (they now trigger a hand-off instead), and adding agent memory to its access and erasure procedures and to its records of processing under Article 30.

Which controls make agent memory compliant?

• A memory policy per agent: what may be stored, why, and for how long.
• Source tags on every remembered fact, so inferences are not mistaken for facts.
• Rights coverage: access and erasure processes that reach every memory store.
• Records of processing that list agent memory as its own activity.
• A data protection impact assessment where memory profiles people or touches sensitive data.
• Separation between customers, so nothing remembered about one leaks into a conversation with another.

The UK and EU regimes start from the same text; differences that matter in practice are covered in UK GDPR vs EU GDPR for AI agents. Firms in Brazil can compare whether the LGPD applies to AI agents. The full map is in AI agent regulation, mapped to controls.