Many negotiations require a counterparty to evaluate information whose disclosure can destroy the value or leverage that made the transaction possible.
NDAI and Conditional Recall
Lets agents reason over private context inside a TEE and reveal only outcomes or memories allowed by explicit conditions.
Key takeaways
Agents can negotiate over private context inside a TEE while revealing only the permitted result.
Precommitting when information may be remembered or discarded makes memory a programmable privacy constraint.
Summary
The talk begins with a recurring information-market problem: one party needs another to assess private information before a deal can proceed, but revealing that information can erase the first party's bargaining position. The speaker illustrated the problem with a public token purchase and a confidential manufacturing recipe.
The speaker connected this pattern to Arrow's information paradox. Traditional non-disclosure agreements address it through legal commitments, but were described as slow, expensive, and imperfect.
NDAI was presented as a technical analogue: each party sends an agent carrying relevant private context into a trusted execution environment, the agents negotiate there, and only a minimal result leaves. The broader idea is programmable privacy, in which information-use constraints are made more specific than a binary choice between disclosure and secrecy.
Conditional Recall asks a related question about memory. It studies what different scenarios permit when participants can define and commit to conditions under which information will be remembered or discarded.
How it works
01
Give each agent private context
Each party gives its own agent the private information needed to evaluate a transaction without sending that context to the other party.
02
Negotiate inside a TEE
The agents negotiate over their private context inside a trusted execution environment, where the raw inputs remain confined.
03
Reveal only the configured outcome
The TEE releases only the agreed minimum, such as a single deal-or-no-deal bit, while keeping the underlying private context inside.
Transcript
I'm not gonna to lie, it's not exactly what I did. But I'll be communicating the vibe of the papers, so that's good. So, I thought Brandon's talk earlier was excellent. And in that talk, he made the claim that, the, like, information markets we exist in are, like, very inefficient. And this line of work is really exploring that same, that same topic. So you can, you can imagine yourself as someone who has private information. You can imagine being in, like, a trading environment where, you want This is actually a MetaDAO is a, is a, like, a futuristic system for investing in DAOs. Not really important, but you can imagine this actually is a problem they have.
Andreessen Horowitz comes, comes to a company and says, "Hey, we will invest a huge amount into this company at, like, a twenty percent discount. Like, do you want to do this deal?" And usually that would be very private information. But in this, like, token-driven world that MetaDAO exists in, all of the DAO, all the people, you need, like, public markets to make this decision. And so Andreessen has to, in public, go and say, "Hey, we want to buy, you know, t-ten million dollars of this token." And everyone else just goes and pumps the token to a super high valuation, like, killing the deal for Andreessen. And so there's this general problem that sometimes you have information, and your counterparty needs to evaluate the information before you have an interaction.
But as you reveal the information to them, you kind of give up, the most valuable part of an interaction. You could also imagine a, factory with, like, a secret sauce for making, I don't know, soda. And they want, they wanna be acquired or they want, to wanna merge with another company. But in order for that other company to evaluate whether it makes sense, they need to see the recipe for the soda. And if you give it to them, then, like, why would they even acquire you in the first place? So this is a problem that comes up everywhere, and actually, there's a name for it in economics called Arrow's information paradox. And in the sort of original, with the traditional world, one way we deal with this is we sign NDAs.
It's, like, a, you know, very expensive process. Takes a lot of time. NDA's aren't perfect, but, okay, our negotiation team and your negotiation team sign NDAs, and then we don't, promise to share information. We promise not to share information about it. And, but what NDAI, this paper I'm about to talk about does, is it captures, the same kind of functionality, but at a technical level using TEEs, and you can use other technology as well. So NDAI is a paper written by Andrew, and that's Stevenson and Shin, and probably someone else who I forgot to mention. And the basic idea is like, okay, well, let's take, these counterparties with private information, and each of them basically sends an agent, with context on their private information into a TEE.
So I'm a company that wants to be acquired. The venture firm, sends their agent. I send mine with data into a TEE. They negotiate, and the TEE is set up so that there's only one bit of information that leaves it, like yes, deal or zero, no deal. And there's no concern about, information leaking. There's no need to involve lawyers, and it's a very expensive and slow process.
And this is, I think, part of a broader, like, commentary on the discussion that happened yesterday, where there was a lot of discussion on the trade-offs around privacy. Oh, do I keep privacy or do I, you know, give up my data for more, a better experience? And technologies like NDAI challenge, this trade-off, and so actually we can use other dimensions and, this concept of programmable privacy, like how making more specific constraints on how your information is used to, create more and more efficient markets without the overhead of things like NDAs.
What is Conditional Recall? Which is another paper you forgot to mention.
Yeah. So I mean, there's a paper about Conditional Recall also by Shin and Christoph Schlegel, from Flashbots. And that is a paper exploring a very similar concept. Like, what is If you can define conditions and commit to conditions under which something is remembered or not, what does that enable? And Conditional Recall explores in a somewhat, literary, somewhat, theoretical, mathematical way, what, how this affects different scenarios. Yeah