I had a brain. The next question was who would use it, and how.
The easiest route is duplication: produce as many agents as needed with the same brain, the same instructions, and the same way of speaking.
That is what I did at first.
I soon saw the limit.
Ten identical agents notice the same things and miss the same things.
It is like sending ten people into a meeting and receiving ten identical sets of notes.
Think of a company, and the answer becomes clearer.
Members of a good organization share the same history and vision, but the finance lead and the sales lead hear different things in the same meeting.
That difference broadens the organization’s field of view.
One meeting, three sets of notes
So I designed the agents to share the entire SSOT while each retained a perspective of its own.
Their assigned roles, what they notice first, accumulated experiences, and habitual ways of judging things all differ.
After the same meeting, the scene looks like this.
Assistant agent
“There are three follow-up actions. One needs a reply by Thursday.”
Research agent
“The market-size figure they gave needs verification. I have formulated two hypotheses.”
Strategy agent
“This proposal conflicts with our direction for this year. Even if we accept it, the terms need to change.”
All three sets of notes came from the same SSOT. All three agents know my values and the company’s context. They are simply standing in different places.
This is what interested me about frameworks such as Letta.
In brief, Letta is an open-source agent framework that grew out of the MemGPT paper by researchers at UC Berkeley.
Its starting point was an analogy with an operating system.
Just as a computer moves between limited memory and a larger disk, an agent moves between memories currently open in front of it and memories filed away.
What matters is that the agent does the organizing itself.
It keeps notes close at hand about “what kind of agent I am” and “who the person I help is.” If it decides while working that a note is wrong, it revises the note itself.
The agent does not disappear when the conversation ends.
It returns the next day with notes a little different from yesterday’s.
If Hindsight in Chapter 1 showed how memory can rework itself, Letta showed that the agent itself could be the owner of that memory.
This starts from somewhere different from making an agent imitate a role through a few lines of system prompt.
An agent accumulates memories of its own over time and develops an identity.
The research agent remembers a hypothesis it made three months ago that proved wrong, and is a little more cautious this time.
These are the scenes I meant in the prologue when I wrote that even the bots were developing a stubborn streak of their own.
The three sets of notes above illustrate role-specific interpretation. Letta documents editable memory and configurations in which agents share memory while retaining their own context. The language of identity and a stubborn streak in this essay is the author’s interpretation of that architecture.
MemGPT - UC Berkeley Sky Computing Lab · Letta memory blocks (V1 SDK) · Letta shared memory
The loop closes
English key to the original diagram. REAL (events in the world) → SSOT → perspectives (each agent’s lens) → a group of agents → actions (emails, meetings, decisions) → REAL. At the center: action becomes reality, and reality becomes the brain again. The green arrow runs from REAL back to the SSOT.
Figure 6. The traces an agent leaves in the world become the agent’s brain again. The green arrow marks where the loop closes.
The point to watch in this diagram is where the arrow comes back.
When an agent sends an email, that email becomes an event in the world. The recipient replies, a meeting is scheduled, a contract moves forward. All of it flows back into REAL and revises the SSOT.
At first, almost all the input circulating through this loop came from me: the people I met, the things I said, the decisions I made.
As the number of agents grows, my share of the data moving through the loop becomes smaller.
That changing proportion is the subject of the next chapter.
A super second brain differs from an army of copies of one brain. It is an organization that interprets the same brain in different ways.
Which different perspectives would you give agents that share the same memory?
This essay is the English translation of Chapter 3 of Super Second Brain, an ongoing Korean whitepaper series by Hanul Lee, with an added source note. Read the Korean series on Substack. Generative AI assisted translation and editing. The diagram comes from the author’s original whitepaper, with an English key above.
