Getting the Brain Right

Last week, we introduced Athena and Archie to the world.

We talked about how we think about AI agents less like chatbots and more like new digital teammates. They have an identity, a body, tools, and a brain. Their identity gives them a way of operating. Their body gives them somewhere to live. Their tools let them take action across the systems we use every day. (You can read last week’s newsletter here.)

But this week, we want to focus on the brain.

Let’s get into it.

The more we work with agents, the more obvious it becomes that memory is one of the biggest differences between a fun AI demo and something that can actually help run parts of a business.

That’s why we’ve been experimenting with building a brain for our agents. 

Instead of treating memory like a static folder of notes, our agent brains now work more like a living knowledge base. It can ingest context from meetings, emails, calls, notes, and conversations, then organize that information into markdown files your agent can return to later. It also builds connections between people, companies, projects, and decisions, creating something closer to a personal knowledge graph.

This matters because most AI memory is still pretty… flimsy. You tell an agent something once, and later it may or may not remember it. Or it only remembers if you ask in almost the exact same words, which isn’t really how humans work. 

The brain tries to solve this with hybrid search. Part of that is semantic search, which means the system retrieves information based on meaning, not just exact phrasing. Under the hood, it turns information into vectors (numerical representations of meaning) so related ideas can be found even when the words are different.

So if we tell Archie, “This client likes short emails with direct next steps,” we don’t want to have to later ask, “What did I say about this client’s preferred email format?”

We want to be able to say, “Can you draft this in the right style for them?” and have Archie pull in the relevant context.

Our brain also combines that semantic layer with keywords and understands how things are connected to each other. That means it can search not only for similar meaning, but also for exact names, entities, relationships, and references. 

As you can imagine, there’s a lot going on in an agent's “brain.” Like humans, agents actually have dream cycles.

Dreaming is the idea that an agent shouldn’t just store everything as raw, messy context forever. While it “sleeps,” it can scan previous conversations, enrich missing entities, fix broken references, consolidate related memories, and summarize loose information into something more structured.

That feels important because people don’t remember by replaying every sentence from every meeting.

We compress. We summarize. We connect dots. We remember that one client cares about speed, that a certain client prefers bullet points, that a project is blocked on one decision, or that a meeting changed the priority of a feature.

Dreaming gives agents a version of that.

Instead of Athena and Archie having a pile of disconnected transcripts, notes, and chat history, they can begin to form more useful working memory: who matters, what changed, what decisions were made, what preferences were expressed, and what context should be retrieved later.

Stay curious,

Julia & Russell