The Future Is for Everyone. But Who Owns the Memory?

Meta’s new essay, “The Future is for Everyone,” begins with one of the right questions about AI: who will have the power?

Its answer is personal superintelligence for everyone.

Meta argues that concentrating advanced AI inside a few companies, governments or institutions would be dangerous. Instead, it wants billions of people to have affordable personal agents, aligned with their goals and available across their work, health, relationships and daily lives.

I agree with more of this argument than some readers might expect.

The AI debate too often offers a choice between catastrophe and automation. Meta’s emphasis on invention and broad access is useful. Its programme providing AI glasses and training to blind veterans is a concrete example of technology increasing independence.

But access to intelligence is not the same as agency.

The difference becomes clearer when we consider the part of a personal agent that may eventually matter most: its memory.

The value is in what the agent knows

Meta imagines an agent that understands you, your goals and everything you care about. It would work continuously across your relationships, health, career, finances and home.

Such an agent would become useful not only because it is intelligent. It would become useful because it has accumulated an understanding of you.

It would know which commitments matter, how you make decisions, what you have tried before, which advice you rejected and why. It would remember the context behind choices that may no longer make sense to anyone else.

Models will improve and may become increasingly interchangeable. That accumulated context will not.

If I use an agent for ten years, its memory could become part of my personal infrastructure. Losing it would not feel like changing a search engine. It could feel like losing part of my history and having to explain myself again from the beginning.

Meta promises strong privacy and a fully private mode. That matters. But privacy is only one part of agency.

Privacy asks: who else can see my context?

Agency asks several additional questions:

• Can I see what the agent believes about me?
• Can I correct something it has misunderstood?
• Can I tell which sources led to a recommendation?
• Can I retire old information that should no longer influence decisions?
• Can I move my history and context to another provider without starting again?

An agent can remember the wrong thing. It can turn an inference into a fact, preserve a temporary preference or give outdated information more authority than it deserves.

A private memory that cannot be inspected, corrected or moved may still be a form of dependence.

Personal agents will enter organisations

The boundary becomes even more important when personal agents participate in organisational work.

An agent acting for an employee needs more than knowledge about that individual. It needs to understand what the organisation has decided, why it was decided, which evidence supports the decision, who has authority and whether the decision is still current.

Suppose an agent helps prepare a board paper. It may draw on private conversations, an old organisational decision, current public evidence and assumptions it inferred from previous work.

Which of those sources may be shared?

Which are authoritative?

Which should be challenged?

Those answers cannot live only inside the private memory of one person’s agent. The organisation needs a shared and governed memory of its own.

When a decision may affect future work, I think an organisation should preserve at least:

• what was decided
• why it was decided
• the evidence or source
• who owns the decision
• who may use it
• when it should be reviewed or retired

Without those distinctions, an AI may find information without knowing whether it is current, authoritative or appropriate to use.

An organisation in which everyone has a brilliant personal agent may still be a forgetful organisation. It may simply produce better answers from fragmented and invisible memories.

Individual empowerment and institutional memory are not opposing ideas. We need both, with clear boundaries between them.

What does “everyone” mean from a small place?

From the Faroe Islands, the word “everyone” carries another question.

A service can be globally available and still be locally absent.

If an AI works brilliantly in English but poorly in Faroese, it is not equally available. If our institutions can use it but cannot govern the knowledge they contribute, they have access without much control. If people must translate their language, culture and public life into the assumptions of a global platform, they are participating on someone else’s terms.

This is not a theoretical concern. UNESCO’s roadmap for multilingualism in the digital era connects language technology with universal access, local participation and data sovereignty.

Small societies do not need to build every frontier model themselves. But they do need the capacity to shape how AI enters their schools, businesses, healthcare, public administration and cultural life.

They need language resources. They need institutions capable of evaluating systems. They need control over their own knowledge. They need the freedom to combine different models and providers without losing their organisational memory each time the technology changes.

A future is not plural because everyone can log into the same service.

Five tests for a future that is actually for everyone

I would therefore add five tests to Meta’s vision.

Useful access. People should have access to capable AI in their own language and circumstances, not merely theoretical availability.

Portable context. A person or organisation should be able to change models and providers without abandoning years of accumulated memory.

Inspectability and correction. People should be able to see what an agent is relying on, correct errors and retire information that should no longer carry authority.

Governed boundaries. Personal, family, organisational, medical and public contexts should not collapse into one undifferentiated profile. Different settings require different permissions and responsibilities.

Participation. Communities and institutions should be able to influence the language support, rules and deployment choices that affect them. They should be shapers of the system, not only its users.

None of this weakens the case for widely distributed intelligence. It describes what meaningful distribution would require.

Distribute the ability to govern

Meta is right that concentration of AI power is one of the defining risks of this period. Affordable access, private modes, open models and independent oversight can all move power in a better direction.

But a provider cannot create a balance of power simply by giving everyone access to a service it still controls.

The future becomes more genuinely shared when people can leave without losing their memory. When organisations can preserve their judgment independently of any model. When small languages and institutions can participate without surrendering the context that makes them distinct.

If intelligence becomes abundant, trustworthy context and legitimate authority may become the scarcer resources.

So yes, distribute the intelligence.

But distribute the ability to govern what it knows, what it remembers and whose purposes it serves.