What's Next

The legendary superbowl commercial where Apple Computer introduced the first Macintosh computer stands atop of what is regarded as one of the most significant ad campaigns ever.  What is notable about that now is certainly the ad, but most impactful is the underlying message that disrupted 30 plus years of the computer industry.  It depicted a dystopian, Orwellian world where a “Big Brother” of the past was overcome by a heroine with eyes on the future.  Are we facing a modern AI dystopian future or are we just crazy enough to think differently?

“You can outsource your thinking, but you can’t outsource your understanding.” — Andrej Karpathy, Sequoia Ascent 2026

Karpathy’s observation captures one of the biggest questions facing artificial intelligence today. AI can now process information faster than any individual and often faster than teams of people working together. But processing information and understanding it are not the same thing. As models evolve from answering prompts to operating as long-running agents, the challenge is no longer simply computation. It is memory: how intelligent systems remember, recall, and reason from experience.

History suggests that when computing enters a new era, it often discovers that the fundamental building blocks have changed. The AI revolution has already experienced this once. It is happening again.

For decades, the CPU was the foundation of modern computing. Computer systems were built around sequential execution because that was precisely what the CPU was designed to do. It became the workhorse of enterprise applications, operating systems, databases, and eventually the internet itself.

Then deep learning arrived.

Researchers discovered that neural networks were fundamentally performing massive matrix operations rather than sequential instruction execution. Fortunately, another processor had been hiding in plain sight for nearly twenty years. Originally designed for rendering graphics, the GPU turned out to be extraordinarily well suited for this new workload.

The breakthrough came in 2012 when Geoffrey Hinton, Alex Krizhevsky, and Ilya Sutskever trained AlexNet using two NVIDIA gaming GPUs, demonstrating a dramatic leap forward in image recognition. The significance of that moment extended far beyond a single research paper. It marked the point at which the AI industry began reorganizing its compute infrastructure around a different architectural primitive.

The GPU did not succeed because it was a better CPU.  It succeeded because it matched the underlying computational primitive of artificial intelligence.

Once that became clear, the entire AI compute stack evolved around GPUs—from training infrastructure and inference systems to networking, memory architectures, and data centers. Importantly, this did not diminish the role of the CPU. Quite the opposite. AI has increased demand across the entire computing stack. CPUs, GPUs, networking, storage, and memory all became more valuable because each contributes something different to modern AI systems.

That observation raises an interesting question.

If AI required a different primitive for computation, could it also require a different primitive for memory?

For decades, computer systems have treated memory primarily as the storage of state. Databases persist the current condition of the world so applications can retrieve it later. That model has served the software industry remarkably well because traditional software is fundamentally deterministic. Developers define the business rules, workflows, and expected query patterns in advance. State provides an efficient representation of the world at a given moment.

Artificial intelligence introduces a different challenge.

As AI evolves from prompt-response interactions toward long-running agents capable of observing, planning, acting, and learning over time, remembering the current state alone is often no longer enough. Intelligent systems increasingly need to remember how they arrived at that state. They need to preserve observations, decisions, tool calls, reasoning steps, feedback, and outcomes so they can explain their actions, evaluate past decisions, and adapt their behavior in the future.

Before exploring what that memory primitive might be, it helps to distinguish between two different ways of storing information.

Think of state as a snapshot and an event as a recording.

A snapshot captures the world exactly as it exists at a particular moment in time. If you want to know the balance of a bank account, the location of a package, or today’s inventory level, a snapshot is exactly what you need. It provides a concise and efficient representation of the present.

A recording tells a different story. Rather than preserving only where you ended up, it preserves how you got there. Every observation, every decision, every interaction, and every change becomes part of a sequence that can later be replayed, inspected, or understood from different perspectives.

That distinction becomes increasingly important for intelligent systems because understanding is rarely built from isolated facts. Humans do not reason by recalling snapshots alone. We reconstruct experiences. We remember what happened, why it happened, what changed, what we tried, what succeeded, what failed, and how one event influenced another. Understanding emerges from that accumulated history.

If one of the ambitions of AI is to approach human-level reasoning, then memory cannot simply be viewed through the traditional lens of RAM, storage, or databases. It must increasingly preserve experience as well as information.

This naturally leads to another question.  

What is the memory primitive of AI?

• The relational database organized around the row.
• The web organized around the request.
• AI memory organizes around the event.

Events capture intent, behavior.

The event is not a new invention. In fact, it has existed for as long as software itself. Every application begins when something happens. A customer places an order. A payment is authorized. A shipment leaves the warehouse. A sensor detects movement. An employee approves a purchase. Something changes in the world, and software responds.

Traditionally, applications consumed those events, executed business logic, updated the current state, and moved on. That architecture made perfect sense because applications already knew which questions they would need to answer in the future. Capturing the current state of the world was usually enough.

Artificial intelligence changes that assumption.

As agents become more autonomous, they increasingly operate in environments where tomorrow’s questions cannot be predicted today. A coding agent may need to reconstruct every decision that introduced a software defect. A cybersecurity agent may need to replay thousands of observations to understand how an intrusion unfolded. A research agent may revisit information it previously ignored because a new hypothesis changes its significance. In each case, the challenge is no longer simply retrieving information. It is reconstructing behavior.

That is why the event becomes very useful.

An event is simply the record that something happened. Unlike state, which represents the world at a particular moment in time, an event naturally captures change as it occurs. It preserves not only what happened, but when it happened, who initiated it, why it happened, how it unfolded, and what happened next. A single event is rarely meaningful by itself. A sequence of events, however, becomes behavior, and behavior is increasingly what intelligent systems need to remember.

This may also explain why the AI industry has struggled to settle on a clear definition for what we often call “memory.” Over the past few years we have heard terms such as prompts, memory layer, context, agent runtimes, observability, harnesses, and most recently loops. At first glance these appear to describe different capabilities. Increasingly, however, they all point toward the same underlying challenge: how intelligent systems accumulate, preserve, and use events over time.

Today, the term context engineering has become one of the more widely adopted ways of describing this challenge. Context is often defined as designing the information environment around the model and it encompasses much more than the prompt itself. It includes memory, planning, evaluation, traces, runtime behavior, tool use, and observations. In other words, context is not simply information. It is an agent’s remembered experience.

This shift reflects a broader change taking place across AI infrastructure. The dominant challenge is moving beyond storing facts toward preserving behavior. Intelligent systems increasingly need to remember not only what they know, but what they have done, why they did it, what they learned, and how those experiences should influence future decisions.  In other words, processing streams of events.

Perhaps the clearest evidence comes from the way modern agents operate.

As Harrison Chase, CEO of Langchain, observed:

“Ambient agents listen to an event stream and act on it accordingly.”

That statement is more significant than it first appears.

We often think of AI systems as consuming prompts, but prompts are only one kind of event. Emails are events. Slack messages are events. Tool calls are events. API responses are events. Database updates are events. Sensor readings are events. Model outputs are events. Human feedback is an event. Agent-to-agent communication is an event. Decisions are events.

Modern agents spend their lives observing, producing, and reacting to streams of events in unlimited and often unpredictable combinations to reason their way to understanding.

Modern AI systems need to remember.

• What happened?
• Why did it happen?
• What changed because of it?

The event provides the natural building block for that memory because it preserves behavior as it unfolds rather than only the state that remains after the behavior is complete.

AI’s new journey

Whenever a new architectural pattern emerges, there is a natural tendency to frame it as a replacement for what came before. History suggests something different.

The relationship between events and state is much the same.

Rather than asking only “What is true now?”, events help answer “How did it become true?”

That distinction is becoming increasingly important because intelligent systems need to explain their reasoning rather than simply produce an answer. They need to reconstruct the observations they made, the tools they invoked, the decisions they took, the feedback they received, and the sequence of actions that ultimately produced a result in a trustworthy, auditable way.

That subtle shift changes the requirements of the persistence layer. Rather than assuming we already know which questions will be asked, AI increasingly rewards architectures that preserve enough behavioral history to answer questions we have not yet imagined.

The history of software has largely been about remembering the current state.

The future of AI is increasingly about remembering the journey and that journey is precisely what events preserve.

They capture not only what happened, but the context that gives those actions meaning. They allow intelligent systems to replay behavior, evaluate decisions, audit outcomes, explain reasoning, and continuously improve over time. State remains the most efficient representation of current reality, while events provide the behavioral history from which that reality emerged.

The GPU became the compute primitive of artificial intelligence because it matched the underlying workload. As AI systems increasingly reason over behavior rather than isolated snapshots, the event plays the same role for memory. It has always been there.  The new AI workload simply reveals why it is now so important.

It’s now the perfect time to be one of the ones who are just crazy enough to think we can change the world!