Toward Agents with a Past

“I want to give Yaya a complete life.”
— Hengyu Tu, The Wandering Earth II

“The past is never dead. It’s not even past.”
— William Faulkner, Requiem for a Nun

Every intelligence has a past.

A human being does not become a self in a single static moment. A newborn enters the world without knowing what danger is, what trust means, what failure feels like, or what is worth pursuing. There is no manual of the world placed in its hands. It learns through hunger, touch, pain, imitation, language, frustration, waiting, and love. Slowly, the world begins to acquire shape.

It remembers which actions bring a response, which choices carry a cost, which experiences deserve repetition, and which mistakes should not be made again. Through continuous perception and action, memory begins to form. Through memory, judgment emerges. Through judgment, preference takes shape. Through preference, something like a stable self begins to appear.

Years later, we call these accumulated traces personality, values, worldview, and eventually, “I.”

This has always fascinated me.

I increasingly believe that intelligence is not merely a computational ability, nor merely a capacity for pattern recognition. Intelligence is a structure formed in time. We become who we are not only because of what we know, but because of what we have lived through. Memory, failure, choice, regret, belief, and the abstractions drawn from experience together shape how we act toward the future.

The past does not mechanically determine the future. Life is not a pre-written trajectory. Yet the past changes the way we imagine what can come next. It influences what we fear, what we trust, what we seek, and what we are willing to abandon. The past is not a passive archive stored somewhere in the mind. It participates in future action.

At a larger scale, civilizations form in the same way.

A family has its memory. An organization has its inertia. A nation has its history. Human civilization is not a static set of rules or a closed textbook. It is a compressed body of collective experience, accumulated across generations through war, migration, cooperation, catastrophe, institutions, technologies, beliefs, and betrayals. We call this compression history.

History is never just a list of past events. It records what humans once believed, what we once misunderstood, how we organized cooperation, and how we reimagined the future after disaster. It tells us which roads once led to prosperity, which fantasies once produced catastrophe, which institutions protected people, and which institutions harmed them. History does not choose the future for us. It shapes the futures we are able to imagine.

From individuals to civilizations, intelligence is always entangled with time. Intelligence needs experience. It needs memory. It needs to accumulate, compress, and abstract from life, then bring those abstractions back into future action.

Today’s AI occupies a strange position.

It has read human history, yet it has almost no history of its own.

It has absorbed massive amounts of text, images, and videos. It can answer questions, generate content, recognize patterns, and simulate reasoning. It can discuss war, science, philosophy, and poetry. It can summarize human experience and predict possible futures. Yet in a deeper sense, it has not truly lived in time. It has no childhood of its own, no body of its own, no continuous experience of acting, failing, waiting, exploring, and regretting in the world.

It possesses vast descriptions of the world, while lacking a world understanding grown from its own experience. It knows many things about what the world is, yet it often does not know how things unfold through action. It can generate a plausible future, yet may fail to connect that future to past failure, physical constraint, and long-horizon consequence. It can imitate the knowledge left behind by humans, while rarely forming an experiential history of its own.

This is one of my deepest questions about current AI:

Without experience accumulated through time, can intelligence become more than imitation?

I do not mean that AI must replicate human emotion, consciousness, or soul. My point is not that intelligence must become human-like. What matters is something more structural: human intelligence is formed through time.

Humans learn not only from information, but from lived interaction with the world. We act, fail, adapt, remember, and gradually revise our internal models through experience. Intelligence, at least in the form we know best, is not static knowledge frozen in a system. It is something that continuously reorganizes itself through feedback across time.

If we want to build intelligent agents that can truly grow, adapt, and revise themselves over long horizons, they may also need to enter time in this way. They need to accumulate experience through perception and action, remember where their predictions failed, and transform those failures into changes in future behavior.

This is why I care about embodied intelligence.

To me, embodied intelligence is not merely a technical subfield. It is an experimental ground for the nature of intelligence itself. Once an agent has a body, or is at least placed inside a world that demands action and feedback, it cannot remain a static pattern recognizer. It must face time. It must act, wait for consequences, encounter resistance, and revise itself through the world’s response.

The physical world is unforgiving. An object will not float because a model thinks it should. An action will not succeed because a plan sounds coherent. If an agent misunderstands gravity, contact, inertia, occlusion, temporal delay, or causality, it will fail. Failure gives intelligence real pressure. Under this pressure, experience stops being mere data and becomes the foundation of learning.

Imagine an agent living in the world.

It may manipulate objects, navigate unfamiliar environments, or interact with other agents whose intentions are only partially visible. At first, it may only recognize patterns. But recognition is not enough. It must learn what remains stable beneath changing appearances, how the world responds to action, and where its own predictions fail against reality.

A failed grasp may reveal something about force and friction. A navigation error may expose misunderstandings of space and memory. A social failure may reveal hidden assumptions about another mind. Across many interactions, the agent should not merely accumulate experiences. It should transform them into structure: an evolving internal model of dynamics, causality, other minds, and its own limitations.

That is the kind of past I care about.

What I want to build is an AI that can accumulate experience in the world.

It may learn from human history, but it should also form a history of its own. It should understand the past, imagine the future, and revise itself through continuous interaction. It should compress experience into structure, abstract local feedback into general principles, and transform past errors into future capabilities.

This is how I understand a world model.

A world model should not be treated merely as a module for predicting the next frame. It is closer to a compressed form of an agent’s experience. It records not only what the world looks like, but how the world changes, how actions alter the world, what kinds of imagination fail, and which structures remain stable across situations. A good world model should function like an internal history of an agent: it grows from the past and participates in future action.

This vision leads me toward three connected directions.

First, an agent must have spatio-temporal experience.

Many current AI systems are strong at static semantics. They can recognize what appears in an image, describe a scene, and align language with vision. But embodied intelligence requires continuous perception-action trajectories. It needs to understand where an object came from, where it is going, why it moves, and what consequences follow an action. Without time, the world is a collection of isolated snapshots. Once time enters the system, snapshots begin to form causality, dynamics, and meaning.

Second, experience must be abstracted.

Humans do not store every experience exactly as it occurred. We learn gravity from countless falls, other minds from countless interactions, and judgment from countless failures. AI needs a similar capacity for abstraction. In the physical world, many structures are not arbitrary. Energy, symmetry, conservation, contact, motion, and object permanence form part of the deep skeleton of the world. If an agent can discover invariant structures and latent dynamics from experience, it begins to move from memory toward understanding.

Third, an agent must recognize its own failures.

A system capable of growth needs to compare imagination with reality: what did I expect to happen, what actually happened, and why are they different? Is the difference merely noise, or does it expose a flaw in my world model? In the grasping example, a failed action is not just a low reward. It is evidence that the agent’s internal prediction did not survive contact with the world. A system that never knows where it is wrong cannot truly improve. For me, the core meaning of a self-improving world model is to give an agent a reflective mechanism over its own experience.

These problems belong to different technical areas on the surface: representation learning, world modeling, physical inductive bias, continual learning, active exploration, and self-improvement. In my mind, they all point to the same larger question:

How can an artificial agent acquire a past of its own?

Research must eventually land in systems, experiments, evidence, and reality. But before research becomes a paper, a model, or a result, it begins somewhere deeper. It begins as a question one cannot let go of.

For me, this is that question.

This path did not begin as a career plan. It entered my life much earlier, almost like a question that had been waiting for me.

When I was in middle school, a copy of Homo Deus mysteriously appeared on my bookshelf. To this day, I still do not know how it got there. I did not buy it. I did not put it there. Later, when I asked my family, they also had no idea why the book was on my shelf. It was simply there, quiet and unexplained, as if an object without origin had entered my life.

When I finished reading it, I felt something close to a shock at the level of the soul. It was not the ordinary feeling of discovering a new interest. It was stronger, stranger, and harder to explain. For the first time, I realized that humanity, intelligence, history, and the future were not distant abstractions. They might be the questions I would have to face with my life.

Perhaps it was coincidence. Perhaps memory gives meaning to accidents after the fact. I cannot prove otherwise, and I do not need to. But in my private mythology, that book has always felt like a sign: not an argument, not evidence, not a doctrine, but a summons. It made the future feel personal. It gave shape to a road before I had the language to describe it.

Since then, my life has carried a clear sense of direction. I have never been able to see research merely as completing tasks, publishing papers, obtaining positions, or entering a system. These things matter, of course. They are part of the world in which research exists. But they are not my true fuel.

I feel as if I have been seized by a question.

What is intelligence? Where is humanity going? What will AI become? Can an artificial agent grow through time and acquire a history of its own?

Perhaps this sounds idealistic. I accept that. But I have never been able to see research merely as solving predefined problems or optimizing within existing systems. What draws me toward embodied intelligence is not its popularity, but the possibility that it forces us to confront intelligence itself: how it forms through experience, becomes grounded in the world, and changes through time.

I believe the most important AI systems of the future will not merely retrieve knowledge or generate outputs. They will perceive, act, fail, remember, revise, and continue evolving through interaction with the world.

That path is difficult. We still do not know how to build agents that can reliably accumulate experience, abstract structure from interaction, and improve without collapsing into error or illusion. But to me, that is exactly why the problem matters.

I want to build agents with a past.

Agents whose intelligence is shaped not only by data, but by what they have experienced, where they have failed, and how they have changed through time.

An intelligence without a past is only a shadow suspended above data.

A true intelligence should leave traces in time.

To build intelligence, give it a past.