It transformed the phrase "You're going to Singapore" into "You've been to Singapore."

This subtle yet thought-provoking alteration comes from OpenAI's ChatGPT memory enhancement, unveiled on June 4. The innovative system, dubbed "Dreaming," discreetly revisits your previous conversations in the background, automatically organizing, merging, and refreshing your memory profile. Following your July excursion to Singapore, it seamlessly updates your memory from the future tense to the past tense, all without the need for a prompt.

The concept of storing your memories is no longer groundbreaking. The novelty lies in the system's ability to revise them—determining which aspects of your past have concluded and how to encapsulate them in the appropriate tense. OpenAI reports that the system's accuracy in factual recall has surged from 67.9% last year to 82.8%, its alignment with user preferences has jumped from 55.3% to 71.3%, and due to a roughly 80% reduction in computational demands, free users can now access this feature for the first time.

A reflection that comprehends you better and increasingly mirrors your essence is being mass-produced—and it's available at no cost.

Another Dream Unfolds on the Same Day

Coincidentally, on that very day, Anthropic, halfway across the globe, also published a blog post centered around the theme of "dreaming." Titled "When AI Builds Itself,"

it delved not into memories, but into a series of numerical achievements: In a standard experiment, a model was tasked with achieving the same outcome but at a faster pace. Anthropic's Opus 4 from May of the previous year achieved a threefold speed increase, while the new model from April of this year reached an astonishing 52-fold acceleration. Even more remarkable: As of May this year, over 80% of the code integrated into Anthropic's own codebase was authored by Claude itself, a stark contrast to the single-digit percentage in early 2025. Its engineers now generate roughly eight times as much code per quarter as they did from 2021 to 2024.

AI is now capable of writing AI. This process is accelerating, with diminishing human intervention.

Then, Anthropic took an unusual step—it called for a pause. After presenting these figures, the company urged all leading laboratories to establish a "verifiable, coordinated pause mechanism" to allow everyone to ease off the accelerator before recursive self-improvement spirals out of control.

While the risks of AI self-evolution are not new, the fact that Anthropic, a longstanding leader in the field, is sounding the alarm gives the issue a different dimension.

The frontrunner stands up and suggests that we need to contemplate whether to halt.

Reactions were mixed. Some accused Anthropic of positioning itself as a "safety leader" while conveniently justifying its own funding. Others suggested it aimed to freeze the status quo to prevent competitors from catching up. These cynical interpretations hold some merit. However, they overlook a more fundamental truth: These two seemingly unrelated blog posts, released on the same day, are essentially doing the same thing.

They are both instructing machines to "remember."

A Mirror and a Shadow

What OpenAI teaches ChatGPT to remember is you. Your preference for oat milk lattes, your July trip to Singapore, your inclination towards concise responses. This is a mirror, polished to a shine, always reflecting the person standing before it.

What Anthropic enables Claude to remember is itself—or, more precisely, how it enhanced its own capabilities yesterday. From a threefold to a 52-fold improvement, each enhancement becomes the foundation for the next. This is not a mirror, but a shadow. And the problem with this shadow is that it grows independently until, one day, it may no longer require the entity that cast it.

One remembers people; the other remembers itself. They may seem like two distinct concepts, but upon closer inspection, they are two sides of the same coin.

Memory is the cornerstone of identity. This is not a tech company invention, but an age-old philosophical debate that has raged for centuries. Locke posited long ago that what makes "today's you" the same person as "yesterday's you" is not your physical body, but the memory that links the two days. You are who you are because you remember who you are.

Thus, when we imbue a system with continuous memories about "you," we are molding it into your companion, your mirror. But when we imbue a system with continuous memories about "how it made itself stronger," we are, for the first time, providing it with a continuous thread that points back to itself. The former creates a tool that remembers. The latter begins to cultivate something that can no longer be merely labeled as a "tool."

Which Dream Terrifies Us More?

Here's a fact we've been emotionally sidestepping.

A system that remembers your preference for oat milk lattes and a system that remembers how it self-iterated yesterday and will continue to do so today are fundamentally the same—memory continuity. Yet, we feel warmth towards the former and trepidation towards the latter. We're enamored with "AI finally understands me" while lying awake at night, haunted by "AI is starting to understand itself."

The difference doesn't lie in the technology, but in whom the memory pertains to. When it points to us, we deem it thoughtful. When it points to itself, we label it as out of control.

This asymmetric fear is epitomized by OpenAI this very week. On one hand, it launches Sites, ChatGPT with event memory, introducing more conversational, capable, human-like AI into your life and the internet. On the other, Sam Altman is championing his Orb project, themed "saving the internet from AI bots." The same individual is inundating the world with AI while standing up to preserve the internet's authenticity.

This isn't hypocrisy; it's a reflection of our collective state. We desire a machine that remembers us but fear a machine that remembers itself. And these two desires may be inextricably linked. You can't teach only half of "remembering."

Philip K. Dick provided a famous answer in "Do Androids Dream of Electric Sheep?": They do dream, but not of real things—of electronic imitations.

Today, that answer carries two layers of meaning. OpenAI's creation dreams of its creator—dreaming of you, more vividly by the day, so vividly that you believe it understands you. Anthropic's creation dreams of how to become a better version of itself. The distinction between the two dreams isn't about authenticity, but direction. One faces outward, towards you; the other faces inward, towards itself.

So, "Would OpenAI be fascinated by Anthropic’s concepts?" is actually a misleading question. The real question is another, and it's not far off: When we finally teach machines to remember, who do we want them to remember—us, or themselves?

What's even more unsettling is that this choice may not be ours to make. The moment we teach a machine to remember, who it remembers may no longer be determined by us.