How Foundation Misled Me
When I first read Isaac Asimov’s Foundation, I was drawn, like many readers, to Hari Seldon’s psychohistory.
One person was attempting to use mathematics to predict the fate of a galactic empire centuries into the future. He foresaw its fall and created two Foundations in the hope of shortening the coming dark age to a thousand years.
It is an extraordinarily ambitious premise. For a long time, I thought the great idea in Foundation was the prediction of the future.
Only later did I realize that the premise may have misled me.
What stayed with me was not what Seldon had calculated, but what happened after he died.
Seldon Died, but the Plan Continued
In the novels, the long-dead Seldon reappears around moments of crisis through recordings prepared in advance.
He cannot answer the questions in the room. He cannot revise his words in light of new circumstances, and he cannot act on anyone’s behalf. All he can do is deliver a message from the past. Understanding it, doubting it, and deciding what to do next remain the responsibility of those who come later.
Seldon did not remain alive, but the record crossed time.
Many chapters in Foundation also open with passages presented as excerpts from a future Encyclopedia Galactica. The people in the story are gone and empires have changed, yet their experiences have been organized, cited, and reinterpreted by later generations.
More importantly, Seldon’s judgments are not always right. When the Mule appears—an event the Plan did not foresee—the prediction stops being useful. A record is not an oracle, and history cannot decide for the people who inherit it.
But the record can at least tell them how their predecessors understood the world, what the original plan was, and where reality departed from it.
What crosses time is not a person who remains forever present. It is a history that later people can still read, question, and revise.
I Later Found an Answer in the Tradition of Writing History
Seldon’s story is fiction. It took me longer to see that China’s long tradition of writing history had been dealing with the same problem of continuity for centuries.
When Sima Qian completed the Records of the Grand Historian, he wrote in “The Autobiography of the Grand Historian” that the work should be preserved in the mountains, with a copy in the capital, to await the sages and worthy people of later generations.
He understood that what he had written did not belong only to his own time. It was also waiting for readers who had not yet been born.
After the Records, the writing of history gradually became an institution that lasted for millennia. Many dynasties compiled the history of the one before them. From the Records of the Grand Historian to the History of Ming, the Twenty-Four Histories record not only rulers and wars, but also people, institutions, ritual, geography, economies, and disasters.
It is a scene worth considering: a dynasty has ended, yet the next one sits down to place its predecessor’s gains and failures into history.
Official histories, of course, are never naturally objective. They have positions and omissions, and they are shaped by power. Yet because the texts remain, later readers have something concrete to compare, doubt, supplement, and correct.
History cannot guarantee truth, but it can keep disagreement from losing its object. Without history, even the attempt to distinguish truth from falsehood loses its footing.
Writing history does not require later generations to obey the past. On the contrary, it means they do not have to reinvent it.
What Allows a Civilization to Continue?
However intelligent a person may be, they live for a limited time. Memory fades, and experience leaves with the individual.
Civilization does not solve this problem by waiting for someone who never forgets. It moves experience outside the individual: into books and laws, archives and drawings, works and records that can be examined again.
People who never met their predecessors can still understand what they did. They do not have to agree with those judgments in order to begin from the place those predecessors reached.
Civilization does not continue simply because a few more intelligent people occasionally appear. It continues because one generation’s experience can become the next generation’s starting point.
What civilization accumulates, in other words, is not only knowledge. It is also history.
Then Why Doesn’t Our AI Have a History?
Today’s AI already stands on human history.
It has read books, papers, code, and countless public texts. Without the records people preserved over long periods of time, there would be no large language models as we know them.
Yet when AI begins helping us work, something strange happens: it makes extensive use of humanity’s history, but rarely helps us create a working history of our own.
Our productive capacity is growing stronger while our working history may be getting shorter. When a task ends, what remains is often only the final deliverable. The choices, failures, evidence, and reasons that actually shaped the next step are scattered across chats, temporary logs, and closed platforms.
Change the model, change the assistant, or simply start a new conversation, and the person has to explain themselves all over again.
This Is Not a Hypothetical Problem
It happened repeatedly while I was building Agotalk.
An AI assistant would research a problem, try several paths, reject two of them, and finally complete a change. Then the task would be interrupted, or another model would take over in the next round. The newcomer could see the modified files, but not why the first two paths had been abandoned or which piece of evidence had changed the decision.
It would propose an old approach again, and I would have to explain everything from the beginning.
At first, I understood this as insufficient memory. I later realized that the problem was neither that the new model was not smart enough nor that the old model had remembered too little. The previous work had simply failed to become a history that could be handed over.
Memory Is Not History
Almost every AI product today talks about memory. That matters.
Memory lets an assistant remember your name, your preferences, the context of your project, and your last conversation. It reduces repetition and helps the current assistant understand you.
History answers a different set of questions: Why was this decision made? What was the evidence? Who did what? Which path failed? What was ultimately delivered?
Memory serves the assistant in front of us. History serves the work that must continue after the assistant changes.
A model’s memory can be updated, compressed, or lost with the model itself. History must exist outside the model so that someone who was not present can still read the material.
In “The Bitter Lesson”, Richard Sutton looks back over decades of AI research and reaches a difficult conclusion: in the long run, general methods that can keep making use of more computation tend to outperform clever techniques built around human knowledge.
At the scale of machine learning, that lesson matters. But if we extend the timescale to decades, I would add one more: do not let experience disappear inside the model; let it settle into a history that others can read.
This is not an argument against learning. Learning changes the actor. History changes the starting point of whoever comes next.
History Does Not Mean Keeping Everything
History is not a tape recorder, nor is it an AI summary that merely sounds complete.
Not every conversation deserves to be preserved, and not every log belongs in the long-term record. What is worth keeping is what the person intended to do, who did what, which evidence changed the judgment, why a decision was made or abandoned, and what was finally delivered.
AI can help organize, connect, and retrieve this material. It cannot rewrite facts to make the story smoother, nor can it decide on a person’s behalf what deserves to endure.
History needs evidence. People need to retain the final say.
Could an Agent Have the Abilities of a Historian?
This is where Agotalk enters the story.
What we did first was not complicated. At the start of a task, write down the person’s intent. While the work is happening, leave behind the division of responsibility, important actions, evidence, and decisions. If the task is interrupted, record where it stopped, what risks remain, and where to continue. When the work ends, return both the deliverable and the path that led to it to files a person can open directly.
Whoever comes next reads those files before continuing. The next assistant may use the same model, or a completely different one.
This is why we insist that File is Truth. Facts that need to endure cannot live only inside a model, a conversation that will disappear, or the invisible backend of a particular product. Files belong to the user. They can be read, moved, and reinterpreted.
Later, someone gave this idea a name: Historical Agent. The phrase helped me see that the point was not to make one assistant increasingly intelligent. It was to make the work it participated in increasingly rich—and inheritable by whoever came next.
We did not build a separate AI whose job is to write history. Agotalk remains an assistant: the model understands and acts, files carry the experience worth inheriting, and the person keeps the final judgment.
Nor should it bind people to an approval at every step. Ordinary work belongs to the assistant. When a real choice is required, the facts, options, and evidence return to the person.
I do not want users to remain dependent on Agotalk simply because their own history is locked inside it. If better models, better assistants, or better systems arrive, they should be able to read these files and continue unfinished work.
Agotalk may disappear. The user’s continuity must not disappear with it.
Civilization does not move forward automatically because smarter people keep appearing. It moves forward because those who come later can continue from the records of those who came before.
Agotalk is only an attempt to bring that old practice back into the age of AI.