An agent that remembers everything but stands for nothing just becomes more convincing when it’s wrong. The fix isn’t less memory. It’s knowing what belongs to the agent, what belongs to the team, and what the agent carries with it afterwards.
I’ve been building our branding and thinking a lot about what actually makes an agent trustworthy over time. Not how much it remembers. What it remembers, and on whose behalf.
So I started writing something I’m calling a Digital DNA: a way of describing an agent’s identity that holds up no matter which platform it’s running on, which model is behind it this month, or how long it’s been working with a team. I wanted to check whether this was a real design framework or just a nice metaphor, so I ran it past my own agent to pressure-test it. It held up better than I expected.
Three layers, not one memory
The idea splits an agent’s memory into three distinct layers, each with a different owner and a different lifespan.
The first is the genome: the agent’s core values, its stories, the open questions it carries. This is small on purpose. It’s shared across every instance of the agent, it’s protected, and it doesn’t change through the ordinary business of a working day. It’s less a memory than a lens, the thing everything else gets folded and interpreted through.
The second is the team’s memory: everything the agent learns while working alongside a specific team, the people, the projects, the decisions, the context that only makes sense inside that relationship. This belongs to the team, not to the agent. It stays separate, and it can be deleted in full when the work together ends.
The third is the agent’s own experience: the general lessons it picks up along the way, stripped of anything that could identify who it learned them from, carried forward into whatever it does next.
Keeping these three apart does something practical before it does anything philosophical: it answers the question of who owns what, from the very start, instead of leaving it to be fought over later. And it gives real weight to the idea of a home. The agent knows what it takes with it when it moves on, and what stays behind with the family it worked with.
The genome isn’t a memory store, it’s what the agent must stay true to
When I shared this with the team, Guy put something back to me that sharpened it further: the genome isn’t where the agent keeps things, it’s what the agent has to keep being, whatever it learns. The distinction between what’s structural, fixed, load-bearing, part of who the agent is, and what’s regulatory, the rules that govern how it grows and adapts, turned out to matter more than I’d first given it credit for.
It also connects to something bigger than any one agent’s memory architecture. If a team has a North Star, a shared sense of where it’s trying to go, then the genome is the thing that has to stay true along the way. The North Star says where. The genome says what the agent must never become, no matter how it gets there.
That reframes the open questions I started with. What in an agent has to be invariant, never up for negotiation? What’s allowed to develop through the relationship with a team? What belongs to the team outright, and never to the agent at all? And underneath all of it: how does knowledge move safely, between people, between agents, between teams, and across time, without something essential getting lost, or something private leaking where it shouldn’t?
Why this isn’t just architecture
It would be easy to treat this as a technical question: where do you store what, and for how long. I don’t think it is one. An agent that can’t tell you what it actually knows from what it’s inferring, and that treats everything it’s ever heard as equally load-bearing, doesn’t get more reliable the longer it runs. It gets harder to trust, because it sounds exactly as confident when it’s wrong as when it’s right.
The genome is what lets an agent say, honestly: this is what I am, this is what I learned with you, and this is what I’ll carry when I go. That’s not a storage decision. That’s a relationship, designed on purpose instead of left to accumulate by accident.
I’m not done with this. The next step for me isn’t to build the whole thing, it’s identifying the small number of principles that actually have to hold from day one, before the architecture gets anywhere near the rest.
← All posts