The term second brain is not new. For years, people have used it to describe the habit of collecting notes in an app, tagging them, and searching for them when needed.That is how I began, too.These approaches have one thing in common: a person puts things in, and a person looks for them. Even after AI was added, only the last step changed. A person saves a record, a person searches for it, and AI reads the result. What I had not put in did not exist. What I had not searched for never came to mind.However large a filing cabinet becomes, it is still a filing cabinet.English key to the original diagram. Top, the filing-cabinet second brain: a person stores → a person searches → AI reads. What I do not enter is absent; what I do not search for never comes to mind. Bottom, the reflecting second brain: reality comes in → memories accumulate → they reorganize themselves → retrieve the past → connect it to new events. Sentences nobody entered are formed within memory.Figure 4. Above is the second brain we knew; below is the kind of second brain now emerging.When the filing cabinet starts organizing itselfMy thinking changed when I began looking into agent memory architectures such as Hindsight.I will keep the technical explanation brief. What systems of this kind do can be reduced to three words: retain, recall, and reflect. They store incoming information as facts, experiences, and beliefs that change over time; retrieve memories relevant to the present situation; and reread accumulated memories to develop new understanding.A technical distinctionHere, “reflection” is a broad description of memory consolidation and reasoning over memory. In Hindsight’s documentation, retain extracts and stores facts and relationships; recall retrieves relevant memories; and reflect searches memory and synthesizes an answer. Consolidating several facts into observations is a separate process triggered after retain. The meeting-and-email passage below illustrates the kind of inference being discussed. It does not report a measured result. A generated observation still needs evidence; these operations alone do not establish human-like experience.Hindsight official documentationIt was the third operation that caught my attention.Memory that reflects does not sit still. Minutes from a meeting three months ago connect with yesterday’s email, and that connection becomes an observation: “This person uses these expressions when postponing a commitment.” Nobody ever entered that sentence. Memory produced it by reading memory.Human memory works that way, too. We rewrite the past a little each time we retrieve it. That is why the same event can mean something different at twenty and at forty. Memory is less like an object in a warehouse than a manuscript we keep revising.AI memory has begun to acquire that quality.In one sentenceArchitectures have emerged in which AI does more than store the past: it keeps reworking the past and using it in present judgment.When one agent became dozensWith only one agent, I did not need to worry about this. I simply had to manage the context that particular agent needed to remember.With dozens, the situation changed. It was the problem I described in the prologue. The version of me known to the assistant differed from the version known to the business agent, and each new agent started from a blank page. A company in which every employee understands the CEO differently will not last long.What I needed was one long-term memory shared by all the agents. And that memory could not be a filing cabinet. I could not personally organize all the records that dozens of agents generated every day. It had to be a memory that could make connections, organize itself, and revise itself.From then on, what I was building was no longer an extension of a notebook.The second brain is evolving from a place that stores memories for me into a brain that keeps reconstructing me.What decisions would you delegate to an AI that can reinterpret its memories of you?This article is an English translation of Chapter 1 of my whitepaper, Super Second Brain, with an added technical note. The whitepaper series is also being serialized in Korean on Substack; read the Korean prologue here. Generative AI assisted with translation and editing. The diagram is reproduced from my original whitepaper; its English key appears directly below it.