AI that remembers customers: what Brazil's LGPD requires when customer service has memory
By Bernardo Gontijo, founder of NeftPublished on Updated on 9 min read
Summary
When customer service AI stores customer history, the company is responsible for that data under Brazil's LGPD: use only what is necessary for the disclosed purpose, limit access and delete it when the purpose ends or the individual has that right, as with data processed on the basis of consent. Deletion is the difficult part: in October 2026, the Bemula benchmark tested six AI-agent memory systems, and none fulfilled more than 1.5% of forgetting requests end to end.
What is customer service AI memory?
Memory is the collection of information from previous conversations that AI consults to provide better service: interaction summaries, open requests and preferences stated by the customer. It is stored outside the language model, in databases and search indexes queried by the AI during each new conversation.
This differs from model training. AI can remember a customer without their data having been used to train anything, and precisely because the memory resides in databases, it can be accessed, corrected and deleted.
For customer service, memory avoids making customers repeat their story at every contact. For data protection, it is another location containing personal data, with the same obligations as any other.
What does the LGPD require of data stored by AI?
Brazil's General Data Protection Law (LGPD, Law No. 13,709/2018) has no specific rule for AI memory: the obligations applicable to any personal-data processing apply. The main ones, applied to memory:
| Requirement | Article | What it means for AI memory |
|---|---|---|
| Purpose | Article 6(I) | Store history to continue service, for the purpose disclosed to the customer, without reusing it for incompatible purposes. |
| Need | Article 6(III) | Remember the useful minimum: a request summary, not the full transcript of every conversation. |
| End of processing | Articles 15 and 16 | Delete when the purpose ends, subject to exceptions such as mandatory legal retention. |
| Data subject rights | Article 18 | Fulfill access and correction requests; on request, delete data processed with consent and unnecessary, excessive or unlawfully processed data. |
| Processing records | Article 37 | Keep records of processing operations, including what enters and leaves memory. |
| Security | Article 46 | Protect against unauthorized access, including access through the AI itself. |
Why is deleting AI memory so difficult?
Because information rarely resides in one place. Something a customer says may appear in a transcript, summary, profile, search index and backups; deleting one copy does not delete the others.
The Bemula benchmark, published on arXiv on October 13, 2026 by Bofan Xie and colleagues, measured this in six AI-agent memory systems, including Mem0, LangMem and A-MEM, using three different language models. The result was negative: no system fulfilled more than 1.5% of forgetting requests end to end, and most fulfilled none. Failures fell into three categories:
- The data is still there: information that should have been deleted could still be extracted using simple questions or prompt attacks.
- Too much is deleted: some systems removed information nobody asked to delete; in one case, the user's address.
- Incomplete correction: when a user replaced a value, the old one survived and reappeared under attack.
“Even after a user explicitly requests deletion, the deleted information remains accessible through search, inference or prompt-based extraction.”
The study has limitations: dialogues are synthetic and only conversation memory was evaluated. External tools, such as databases and logs, were excluded and are additional places where data may persist.
What if the law requires keeping what the customer asked to delete?
In that case, the legal obligation takes precedence. The LGPD allows data retention to comply with legal or regulatory obligations (Article 16(I)), and the right to delete data processed with consent makes an exception for these situations (Article 18(VI)).
Retention also depends on sector regulation. For covered providers, Brazilian Decree No. 13,135/2026 requires history for at least six months after resolution, and telephone recordings and voice or text messages for at least 90 days after service (Articles 18–19). It replaced Decree No. 11,034/2022. Separate two things:
- Service record: the reference number, transcript and recording required by law or contract. They remain for the required period with restricted access, then are deleted.
- AI working memory: summaries and preferences the AI consults to serve the customer next time. Without a retention obligation, the company can delete them on request; if memory depends on the customer's consent, deletion is their right (Article 18(VI)).
When a request cannot be fulfilled immediately, the company must explain the reason to the data subject (Article 18, paragraph 4). Requests must be fulfilled at no cost to the individual (Article 18, paragraph 5).
Who can see what AI remembers?
Only those authorized to see the original data, and the rule must apply within search, not after it. This was the central point of a talk by Aaron Linsky (Bridgewater Associates) and Nikhil Benesch (turbopuffer) at AI Engineer in October 2026, about an AI assistant searching 100 million internal documents.
At Bridgewater, each document has different permissions by person, group and licensed source. The speakers show that the naive approach—retrieving the most similar documents and then discarding those the user cannot see—loses relevant results and becomes slow at scale. Their solution was to apply permissions within search itself.
In customer service, the same logic translates into three rules:
- Identity before history: AI consults previous interactions only after verifying it is speaking with the data subject.
- Role-based permission: agents, supervisors and AI itself see only what their role requires.
- Customer text is not an instruction: what the customer writes, or what appears in a document they submit, is recorded as a statement, never as a rule for the AI.
The third rule addresses what research calls authority laundering: external content stored in memory comes to be treated as a legitimate preference or order. SAFE-Mem, published on arXiv in October 2026, proposes labeling every memory item with its type, the authority of its producer and permitted scope of use, then filtering by these labels before AI uses it. The study's tests used controlled scenarios, not production deployments.
What does the research change in day-to-day service?
| What the source shows | What changes in customer service | Limitation |
|---|---|---|
| Bemula: no system fulfilled more than 1.5% of forgetting requests | Test deletion end to end, across all copies | Synthetic dialogues; conversation memory only |
| Bridgewater: permissions applied within search | History only after identity verification and according to role | Production report with a commercial interest from the provider |
| SAFE-Mem: type, authority and scope labels in memory | Separate customer statements from operational rules | Controlled scenarios, no production use |
What should you ask an AI provider before buying?
Ask for written answers, preferably demonstrated, to these questions:
- Where are recordings, transcripts, summaries and AI memory stored, and in which country are they processed?
- Can retention periods be configured for each data type, and what happens when they expire?
- How is a deletion request carried out across all copies, including search indexes, logs and backups? Is deletion recorded?
- Does AI consult customer history before or after verifying identity?
- Who can consult the history, and is access logged?
- Is interaction data used to train models? Under what lawful basis, and how can you refuse?
- How does AI handle instructions written by customers or present in submitted documents?
- Which subprocessors, such as AI model providers, receive data?
How does Neft handle data in AI customer service?
In Neft's AI customer service, AI works in the same queue and under the same reference number as the team, and the reference number, recording, transcript and reports stay in the same service record. The virtual agent only queries systems or performs actions authorized by the organization, with identity verification when configured; text instructions do not expand those permissions.
AI service audio is processed by Neft's servers and Neft's contracted AI provider. For the rest of video operations, see the LGPD guide for video customer service.
Frequently asked questions
- Does the LGPD prohibit AI from remembering previous customer conversations?
- No. The LGPD requires a lawful basis, a disclosed purpose, data minimization and security. AI may use history to continue an interaction, provided the customer is informed and memory complies with these rules.
- Can a customer ask AI to forget what they said?
- They can request it, and in many cases the company must comply. Article 18 of the LGPD guarantees deletion of data processed with consent and of unnecessary, excessive or unlawfully processed data. Exceptions include data the company must retain by law, such as service records for regulated customer service.
- Does deleting data from the database guarantee that AI has forgotten it?
- Not necessarily. The Bemula benchmark (October 2026) showed that AI memory systems leave information extractable or delete beyond the request. Deletion must reach summaries, search indexes, logs and backups.
- How long can I retain a service transcript?
- The LGPD sets no single period. For providers covered by Brazilian Decree No. 13,135/2026, Articles 18–19 distinguish history (at least six months after resolution) from telephone recordings and voice or text messages (at least 90 days after service). Validate transcript classification, purpose and sector rules with legal counsel.
- Can I use service recordings to train AI?
- Only with a lawful basis and purpose compatible with what the customer was told (LGPD Article 6(I)). Ask the provider whether it uses your service data to train models and how to refuse.
Sources
Videos, studies and documents used in this article. Preprints are versions published by their authors before peer review. Article researched and written with AI assistance; research and quotes checked on October 26, 2026. Customer service retention updated on November 16, 2026 against Decree No. 13,135/2026 on Brazil's official Planalto website. This is not legal advice.
Research paper (preprint)
Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents (opens in a new tab)Bofan Xie, Chaokun Wang and others (arXiv 2610.18052, version 1) · published on
None of the six evaluated memory systems fulfilled more than 1.5% of forgetting requests end to end; synthetic dialogues.
Research paper (preprint)
SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval (opens in a new tab)Haoming Li, Bingzhe Li and others (arXiv 2610.22959, version 2) · published on
Type, authority and scope labels in memory to prevent authority laundering; tests in controlled scenarios.
Video
Can Your AI Analyst Keep a Secret? (opens in a new tab)Aaron Linsky (Bridgewater Associates) and Nikhil Benesch (turbopuffer), AI Engineer channel · published on
Permissions applied within search across 100 million documents. turbopuffer is the provider of the presented solution.
Legislation
Brazilian General Personal Data Protection Law (Law No. 13,709/2018) (opens in a new tab)Office of the President of Brazil · published on
Purpose and necessity principles, end of processing, data subject rights, records and security.
Legislation
Brazilian Decree No. 13,135 of October 9, 2026 (customer service regulation) (opens in a new tab)Office of the President of Brazil · published on
Retention of history, telephone recordings and messages; Articles 18–19.
Read more
Want to serve your customers by video?
Take the free assessment or try Neft for 15 days, with no credit card.


