Superintelligence and LGPD: Keeping Agent Memory Lawful in Brazil
Brazilian firms can keep AI agent memory within LGPD by setting purpose limits, answering data subject requests and enforcing retention, step by step.

Superintelligence and LGPD is a practical question more than a separate legal regime: can a company let a capable AI agent remember customer and employee details without breaking the Brazilian data protection law's rules on purpose, necessity, data subject rights and retention? For most Brazilian firms the answer is yes, but only when agent memory is designed as a governed data store with a stated purpose, a named owner, a request process and a deletion schedule, rather than an open notebook the agent keeps indefinitely.
Key takeaways
- Treat agent memory as personal data processing whenever it holds information about identifiable people, because in most cases it does.
- Write down one purpose for each memory store before the agent writes anything into it.
- Make data subject requests reach memory, logs, summaries and backups, not only the chat interface people see.
- Set retention by purpose, delete on a fixed schedule, and keep a short record of what was removed and when.
- Keep a named person in the approval path for memory changes that touch sensitive data or new sources.
What the LGPD asks of agent memory
The Lei Geral de Proteção de Dados Pessoais does not mention AI agents, and it does not need to. Its principles apply to any processing of personal data, and an agent that stores what a customer said, what an employee asked for or what a supplier's contact prefers is processing personal data in the ordinary legal sense. The principles that matter most here are purpose, necessity, adequacy, transparency, security and accountability, all listed in the law's sixth article.
The practical consequence is that memory cannot be treated as a technical convenience. A model that remembers a user's preferred name is not a neutral feature. It is a store with a reason to exist, a set of people whose information sits in it, and an operator who must be able to explain both. Companies that already keep a record of processing activities can add agent memory as a new entry. Companies that do not yet have one can start with the memory store, because it is usually the least understood place where personal data accumulates.
The controller remains responsible even when a vendor hosts the model. That is why contract terms, subprocessors and data locations matter alongside the design of the memory itself. The LGPD and AI agents checklist for Brazilian companies covers the questions to ask before an agent touches personal data for the first time.
Purpose limits: write down why each memory exists
Purpose limitation is where most agent memory problems begin. Teams often switch on memory because the product offers it, then discover months later that the same store holds support history, sales notes, hiring comments and internal debugging traces. Each of those carries a different expectation from the people involved, and a single store cannot serve all of them lawfully without deliberate boundaries.
A workable rule is one memory store per purpose, with a written statement of what the store is for, what it may contain and what it must never contain. The statement should be short enough that a manager can read it in a minute and specific enough that a developer can enforce it. Phrases such as "improve customer experience" are too broad to limit anything. Phrases such as "remember the open order status and the agreed delivery window for a customer during an active support case" can be checked against what the agent actually writes.
The agent should also be unable to widen its own purpose. Memory writes should pass through a policy layer that rejects categories outside the stated scope, such as health information, national identification numbers or private family details that appear in a free-text message. This is not only a legal safeguard. It keeps the store small, which makes every later step easier.
A memory store without a written purpose is a collection waiting for a reason.
Illustrative scenario: a support agent's notes
The following scenario is illustrative and does not describe a real company. A regional furniture retailer deploys an agent to answer delivery questions. The agent is allowed to remember the order number, the delivery window and the status of an open claim for the length of that case. A customer mentions in passing that a relative has a medical condition and asks for a delivery slot that fits the relative's treatment schedule. The policy layer stores only the delivery window, discards the medical detail from memory, and flags the conversation for a person to handle the sensitive note under the company's own procedure. The retailer's operations lead reviews a weekly sample of flagged conversations to confirm the rule works in practice.
Data subject requests must reach the memory
The LGPD gives people rights over their personal data, including confirmation that processing occurs, access, correction, anonymization or deletion of unnecessary data, portability, information about sharing and revocation of consent where consent was the basis. An agent that remembers information creates a new place where those requests must be answered. If a person asks for deletion and the company removes the account record but leaves the agent's memory untouched, the request has not been fulfilled.
The operational answer is a request path that names every location where a person's information may sit: the primary memory store, session summaries, vector indexes built from conversations, logs, analytics exports and backups. Each location needs an owner and a method for locating a person's records. Search by a single identifier is rarely enough, because people refer to themselves by nicknames, phone numbers that change, or by describing an order rather than giving a name. The guide on deleting agent memory on request walks through how to find, remove and verify that removal in practice.
Verification matters as much as removal. A deletion record should show which stores were searched, what was removed, what could not be removed and why, and when the check was performed. Where backups cannot be altered immediately, the record should say so, and the company's procedure should explain how those copies expire or are isolated. The ANPD's own publications at [the national authority's portal](https://www.gov.br/anpd/pt-br) are the right place to confirm current expectations on response handling, which can change as guidance develops.
Retention: decide how long memory lives
Retention is the question that agent designs most often leave open. Models and platforms tend to store context by default, and an agent with a long-running task may carry information far beyond the reason it was collected. The LGPD expects personal data to be eliminated when its purpose has ended, subject to specific exceptions such as legal obligations, so every memory store needs an end point defined in advance.
A useful retention schedule has three parts. The first is a lifetime for each purpose, measured from the event that closes it, such as the end of a support case or the end of a contract. The second is a trigger that starts deletion automatically, rather than relying on someone to remember. The third is an exception list, short and approved, for information that must be kept because another law or a legal hold requires it. Keeping that exception list visible prevents informal retention from appearing quietly across the system.
Summaries deserve special attention. Many agents compress long histories into summaries, and the summary often outlives the detailed record it came from. If the detailed record is deleted but the summary keeps a sensitive fact, the deletion has been incomplete. Summaries should inherit the retention schedule of their source, or be regenerated after deletion with the removed material excluded.
A step-by-step plan for the first memory store
Brazilian firms that want to start lawfully rarely need a large program. A single agent with one memory store, handled carefully, teaches most of what a larger rollout would require. The following sequence is a practical starting point, and each step produces a document or a setting that can be reviewed.
- Name the single business task the agent performs and list the personal data it genuinely needs to complete that task.
- Write the purpose statement for the memory store, including what it must never contain.
- Map every location where the agent's inputs, outputs, summaries, logs and backups are kept, and name an owner for each.
- Set a retention period for each purpose and an automatic deletion trigger, with a short list of approved exceptions.
- Build a request path for access, correction and deletion that searches every mapped location and records what was done.
- Add a policy check that blocks writes of sensitive categories and routes borderline cases to a person.
- Run a test with fabricated records to confirm that deletion removes the data from memory, summaries and indexes.
- Review a sample of stored memories on a fixed schedule and record the findings, changes and the reviewer's name.
The last step matters most over time. Memory drifts. Prompts change, new integrations are added, and staff begin using the agent for tasks nobody planned. A regular sample review catches drift early, and the record of those reviews is also evidence of accountability if the company is ever asked how it governs personal data.
How Be Human handles this in managed operations
Be Human designs and operates personalized AI infrastructure in which an agent's identity, memory, channels, tools, permissions and operating context stay under the client's control. For memory and LGPD, that means each memory store is defined with a purpose and retention schedule before the agent is connected to live data, deletion requests are routed through the same mapped locations every time, and writes of sensitive categories are blocked or escalated to a named person. Managed operations add recurring permission reviews, sampled checks of stored memory, and documented incident response when something is written that should not have been. The first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, so the company can verify the setup before expanding it.
If your team is weighing which agent task deserves memory first, the free Be Human AI diagnostic will help you identify the first task worth handing to an agent.
Frequently asked questions
Does the LGPD apply if our agent only stores notes about internal staff? Yes, in most cases it does, because information about identifiable employees is personal data in the ordinary sense. The principles of purpose, necessity and security still apply, and employees' data subject rights apply as well. The safest approach is to treat staff memory with the same mapping, retention and request handling used for customers.
Can we keep agent memory indefinitely because the model needs context to work well? Keeping context indefinitely is rarely justified by the LGPD's principles, since personal data is generally to be eliminated once its purpose has ended. Good performance can usually be preserved with shorter, purpose-bound memory and well-designed summaries. Where a longer retention period is necessary, the reason should be written down, approved and reviewed.
Is deleting data from the chat interface enough to answer a deletion request? No. A deletion request must reach every place where the information is stored, including memory stores, summaries, vector indexes, logs, exports and backups that the company controls. The company should also keep a record showing what was searched, removed and verified, and explain any copies that cannot be changed immediately.
What should we ask our AI vendor before letting an agent store customer information? Ask where memory is stored, which subprocessors can access it, how long the vendor retains inputs and outputs, and how deletion is carried out and verified. Ask whether the vendor's terms allow the data to be used to train other systems. The answers should be reflected in your data processing agreement, and the vendor's documentation should match what the contract promises.
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.


