This article is general information, not legal advice. It describes design questions, not legal conclusions. I am not a lawyer, and whether a particular memory design complies with the Protection of Personal Information Act (POPIA) depends on the organisation's circumstances. Those determinations belong with your legal and privacy people. Read the Act itself, and take advice.
With that said, there is a practical gap worth closing. Privacy reviews of AI assistants tend to focus on where the model runs and whether the provider trains on submitted data. Those are real questions. But an assistant that remembers customers creates something more durable than a model call: a store of personal information about people, built up over time, partly from what they said and partly from what software concluded about them. That store is where most of the lasting obligations sit.
This part translates POPIA's conditions into design questions for agent memory, using Sunbird's support assistant. It follows the series' running example rather than any real organisation's arrangements.
Memory is processing of personal information
Sunbird Home is fictional, and its systems and figures are illustrative. Every item Sunbird Help remembers about Zanele Mokoena is personal information under POPIA. That is obvious for her name, phone number and order history. It is less obvious, and just as true, for two kinds of item memory systems tend to accumulate.
POPIA's definition of personal information includes "the personal opinions, views or preferences of the person". Her preference for WhatsApp updates is personal information. The definition also includes "the views or opinions of another individual about the person". A stored note that a customer "seemed angry" or "is likely to cancel" is an opinion about her, written by software, and it is personal information too. This is one of the reasons Part 4 kept inferences out of memory.
There is still no AI-specific statute in South Africa. A draft national AI policy published for comment earlier this year was withdrawn on 26 April 2026. Memory in an AI system is processed under POPIA in the same way as any other personal information, which is useful: the conditions are known, and most organisations already have people who work with them.
The eight conditions, asked of a memory store
POPIA sets out eight conditions for the lawful processing of personal information. Each turns into a concrete question for a memory design.
Some of these need more than a line.
Accountability (section 8). Someone must own the memory store, its rules and its behaviour. At Sunbird that is the head of customer experience, with the platform team as custodian. "The vendor's memory module handles it" is not an owner.
Minimality (section 10). Personal information may only be processed if, given the purpose, it is adequate, relevant and not excessive. For memory, this is the write policy's first question: what is this worth remembering for? Most conversational detail fails it, and should never be stored.
Retention (section 14). Records must not be kept longer than necessary for the purpose they were collected for, subject to listed exceptions such as retention required by law or by contract. When an organisation is no longer authorised to keep a record, section 14(4) requires it to destroy, delete or de-identify it "as soon as reasonably practicable". A memory store with no expiry is, by design, not doing this.
Further processing (section 15). Further processing must be compatible with the purpose of collection. A preference collected to send delivery updates is a poor basis for a marketing campaign or a credit decision. Sunbird's rule is that support memory is not reused for other purposes until someone has checked, and the default answer is no.
Information quality (section 16). Reasonably practicable steps must be taken to keep personal information complete, accurate, not misleading and up to date where necessary. Copying system-of-record facts into memory works directly against this, because copies go stale. So does an assistant that keeps using a preference nobody has confirmed in years.
Openness (sections 17 and 18). People should know their information is being collected. If Sunbird Help remembers preferences between conversations, customers should be told, plainly, at the point where it happens and in the privacy notice.
Security safeguards (sections 19 to 22). Appropriate, reasonable technical and organisational measures are required, and any operator processing on Sunbird's behalf must do so under a written contract that requires those measures. For memory, the operators may include the model provider, a memory or vector-database service and the helpdesk platform. Each should be identified and contracted. Compromises must be notified under section 22.
Data subject participation (sections 23 to 25). People may ask what personal information is held about them, and section 24 lets them ask for information that is inaccurate, irrelevant, excessive, out of date, incomplete, misleading or unlawfully obtained to be corrected or deleted. A memory design must make that possible in practice, which means being able to find every item about a person, including derived ones.
Special personal information: keep the instruction, not the reason
Section 26 starts from a prohibition on processing special personal information, which includes information about a person's health, religious or philosophical beliefs, race or ethnic origin, trade union membership, political persuasion, sex life, biometric information and criminal behaviour. Exceptions exist, but they are specific.
Customers volunteer this kind of information in support conversations all the time, usually to explain a request. Zanele once wrote: "Please make sure they deliver to the ground floor, I use a wheelchair."
Sunbird's design keeps the instruction ("ground-floor delivery") as a delivery preference and does not store the reason. The courier needs to know where to deliver. Nobody at Sunbird needs a record of the customer's health to do that. The same pattern applies widely: when a customer explains a request with sensitive context, store the request and drop the context.
Children's personal information is a separate matter. It is not part of section 26. It is dealt with in sections 34 and 35, which have their own prohibition and authorisations. Sunbird's customers are adults, but parents sometimes mention children in passing ("my daughter will sign for it, she's 15"). The write policy treats any detail about a child as something not to store.
Forgetting is a feature
The most common gap in agent memory designs is not what goes in. It is what never comes out. A memory store without a deletion policy keeps growing, gets more personal and becomes harder to search, correct or explain. It turns into a liability that grows quietly.
Deletion is also harder than it looks, because a single remembered item is rarely in one place.
When Zanele asked Sunbird to forget her WhatsApp preference, the row in the preference store was the easy part. The same preference had been mentioned in a closed case record, which was embedded in the helpdesk's search index. It appeared in a conversation summary and in application traces. One of her conversations had been copied into an evaluation set. And all of it was in the nightly backups.
Sunbird's deletion process now covers every one of those places. Live stores are updated at once. Summaries and case-search embeddings are regenerated without the item. Traces are redacted on their normal expiry. Evaluation cases built from real conversations are removed or replaced with synthetic ones. Backups age out on a stated cycle, 35 days in Sunbird's case, and are not restored into live systems without the deletion log being reapplied.
None of that is possible unless the write path recorded where each item went. Deletion is designed at write time.
Questions to take to your privacy team
Bring these to the people who will make the legal determination, with your design in hand:
- What is the stated purpose of each kind of memory, and is it written down?
- Which items are personal information, and which might be special personal information?
- What is the retention period for each kind, and what triggers deletion?
- Who are the operators, and do the contracts cover memory, embeddings, logs and backups?
- Is any memory transferred outside South Africa, and on what basis? CloudNala's plain-language article on why POPIA does not require your data to stay here explains section 72's conditions.
- How does a customer see, correct or delete what is remembered, and how long does it take?
For a broader starting point on which data can safely go into AI at all, see the traffic-light approach.
Key points
Memory is personal information. Including preferences and opinions about people, which POPIA's definition names.
The conditions become design questions. Owner, minimality, purpose, retention, quality, openness, security and participation.
Keep the instruction, not the reason. Sensitive context behind a request rarely needs storing.
Children's data sits outside section 26. It has its own sections, 34 and 35.
Deletion is designed at write time. One remembered item can live in six places.
Next: Part 7, Evaluate agent memory as behaviour. All parts are on the series page.
How CloudNala can help
We help teams prepare a memory design for privacy review: what is stored and why, where every copy lives, the operators involved, the retention and deletion path, and how a customer's request is handled end to end. We do not give legal opinions. We give your legal and privacy people a design they can assess.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an Agent Memory Review or write to us at consult@cloudnala.co.za