Latest / B2B SaaS Talks with Fexingo: Enterprise Software, Sales Cycles, and Procurement / Enterprise RFPs Now Ask About Vendor AI Agent Memory
Transcript
- Lucas: So I was reading through a recent RFP from a large healthcare network — this is a real one that crossed my desk — and buried in the security section, there was a question I hadn't seen before. It asked about our AI agents' memory: how we store it, how we segment it per client, and what happens to that memory when the contract ends. Luna: Memory. That's a big step beyond just data residency or model rollback, which we've covered before. What exactly are they asking for? Lucas: Right. So memory in the AI sense is the context the model carries across sessions. Think of a support agent that remembers your company's specific policies, or a sales copilot that recalls the last conversation it had with a prospect. The RFP wanted to know if that memory is per-user, per-tenant, and whether it could be exported or deleted on demand. Luna: Export and delete. That's essentially gdpr style rights, but applied to AI context. Makes sense for a hospital network — they have to be careful with protected health information. Lucas: Exactly. And this isn't a one-off. I've seen similar questions in at least three other RFPs this quarter, all from regulated industries — finance, healthcare, education. The pattern is clear: buyers are starting to treat AI memory as a data asset that needs governance. Lucas: But here's the tricky part. Most vendors haven't built memory with that in mind. They've treated it as a technical feature, not a compliance obligation. And that's a gap that's going to become very visible in the next twelve to eighteen months. Luna: Is that because the technology is still evolving? I mean, memory is a relatively new capability for AI agents. Lucas: Partly. But it's also because the enterprise procurement cycles lag the product development. Vendors ship memory because it makes their assistant feel smarter, but they don't think about the legal and security implications until a buyer asks. Luna: So what does a vendor need to answer that question convincingly? What are the concrete things they need to have in place? Lucas: First, you need to know exactly what your system remembers. That sounds obvious, but a lot of agents use a vector database that stores embeddings — essentially compressed versions of the conversation. If you can't explain what's in there, you can't attest to anything. Lucas: Second, you need to be able to isolate memory per customer. That's a technical architecture choice. If your memory is shared across customers, even in a pooled model, that's a red flag for enterprise buyers. Luna: So tenant isolation is the baseline. Then what? Export and deletion? That's a feature, not just a policy. Lucas: Exactly. And that's where a lot of vendors are scrambling. You can't just say 'we delete data when you ask' if your memory is embedded in a model's weights. You need to be able to remove the specific vectors or records that pertain to that customer, and do it in a defensible way. Lucas: One thing I've seen vendors do is build a memory management API. It lets admins view, export, and delete memory entries. That's the kind of thing that makes procurement teams feel comfortable. Luna: And it also helps with your own compliance, right? I mean, if you can't delete a customer's memory, you're basically holding their data hostage. Lucas: Right. And that's a liability. Especially if they move to a competitor. They'll want their data back. Luna: So let's talk about the sales side. How does this change the conversation with a buyer? It's not just a technical checkbox. Lucas: It's become a trust signal. When a vendor can articulate their memory governance, it shows they understand the enterprise stakes. It's a differentiator. Lucas: And on the flip side, if you can't answer, you might lose the deal. Procurement teams are adding this to their evaluation criteria, even if it's not an explicit RFP question. They'll ask in the demo: 'What happens to our data in your AI assistant?' Luna: So it's like the new SOC 2. In the past, buyers wanted to see your security certification. Now they want to see your AI governance. Lucas: Exactly. And that's a big opportunity for vendors who get ahead of it. If you have a clear memory policy, you can publish it, make it part of your security docs, and use it to close deals faster. Lucas: But here's the thing — a lot of founders I talk to are intimidated by this. They think it's a technical deep dive. But it doesn't have to be. Start by mapping what your memory does, and then build the controls around it. Luna: That sounds like a lot of work, but given how many RFPs are now asking, it might be worth it. I mean, we've seen the adoption curve on these things — it's not going away. Lucas: And that's exactly why we're doing this episode. If these conversations are useful for what you're building or running, you know, a couple of dollars a month is genuinely what keeps these going — buy me a coffee dot com slash fexingo, if you've gotten something out of them. Luna: Yeah, it's a small thing that makes a real difference for us. Lucas: So, back to the practicalities. Let's talk about how a vendor should actually approach this. Where does memory fit in their compliance stack? Lucas: I'd say it belongs in the data processing agreement, right alongside data retention and deletion policies. You should have a specific section on AI agents and their memory. Luna: That makes sense. It's a type of personal data, after all. And for some industries, it's protected health information or financial data. Lucas: Right. So the first step is to classify the data your agents touch. If your AI assistant is pulling in customer records, that's sensitive. You need to know what's stored in memory and how long it lives. Lucas: Then you need to set retention limits. Some vendors are moving to short-term memory — keeping only the most recent context and forgetting the rest. That reduces the attack surface and makes compliance easier. Luna: Short-term memory. Interesting. So you don't need to remember everything forever. Just enough to be useful in a session. Lucas: Exactly. And that's a design choice that can simplify things a lot. Instead of building complex deletion machinery, you just don't store long-term memory in the first place. Lucas: But some use cases need long-term memory — like a sales copilot that needs to remember a prospect's preferences from months ago. In that case, you need a clear lifecycle: how it's created, how it's accessed, how it's deleted. Luna: And I assume that's where the export feature comes in. For a customer who wants to move to another vendor, they need to be able to take their data with them. Lucas: Or at least verify that it's been deleted. In the healthcare RFP I mentioned, they wanted a certificate of deletion within thirty days of contract termination. That's a specific requirement that most vendors can't meet today. Luna: So that's a huge opportunity for a vendor to win a deal if they can actually do that. Lucas: Absolutely. And we're seeing some startups build their entire product around memory governance. They offer an audit log of what the AI remembers, and they can prove compliance. Lucas: But for established SaaS companies, it's about retrofitting. They have to go back and add these controls to their existing agents. It's not trivial, but it's doable. Luna: What about the AI ethics side? Does memory raise new ethical questions beyond just compliance? Lucas: It does. For example, if an AI remembers something a user said in a moment of anger, and then uses that against them later, that's a problem. You need to think about what kind of memory is appropriate. Luna: So it's not just about what's legal, but what's fair and transparent. Lucas: Right. And that's why some companies are adding a 'memory policy' to their AI governance committee — which we covered a few episodes ago. They're setting rules for what types of information can be stored and for how long. Lucas: So if you're a vendor, I'd say start by doing an audit of your AI features. Ask yourself: what do we remember, and can we prove it? Luna: That's a good place to start. And I'd add: check your existing RFPs and see if you've already received questions about memory. If you have, that's your signal. Lucas: And if you haven't, they're coming. The fact that we're seeing this in healthcare and finance now means it'll spread to every vertical within a year. Lucas: I'm actually excited about this. It's a chance for vendors to differentiate on trust, and for the industry to mature a little. The vendors that treat memory as a feature will get left behind, but the ones that treat it as a responsibility will win. Luna: That's a great note to end on. Thanks for walking us through this. Lucas: Thanks for having me. And if you're building in this space, I'd love to hear how you're handling memory. We're on LinkedIn.