I keep coming back to one problem with AI subscriptions:

Every time I switch models, I lose a part of the work I have done with AI.

I use different models for different reasons. ChatGPT, Claude, Kimi, Sakana, Grok, coding agents, whatever is particularly good at the thing I need at that moment.

Sometimes a vendor releases a feature I like just as my subscription somewhere else expires, so I move.

That should be a good thing.

We should be able to switch AI vendors easily. Competition between models is moving incredibly fast, and I do not think anyone should have to marry themselves to one provider.

But switching has a strange cost.

During the month I spend with a model, life happens.

I might spend a week deep in a technical problem. Then another researching an industry. Then debugging an architecture, thinking through a business decision, studying a regulation, writing something, or learning an entirely new domain.

During that process I am not just consuming inference.

I am building context.

The AI learns enough about the problem to become useful. I learn from the AI. We explore dead ends. We discover better approaches. I explain why certain decisions were made. It sees the architecture of a project, the assumptions behind it, what failed before, how I prefer to work, and what I eventually learned.

By the end of that process there is something valuable between us that did not exist at the beginning.

Then I change AI providers.

And suddenly I am introducing myself again.

That feels wrong.

The important asset is not the raw chat history and it certainly is not the tokens themselves.

It is the accumulated state created through the interaction:

  • the decisions
  • the research
  • the connections between ideas
  • the things that failed
  • the things that worked
  • the project context
  • the workflows
  • the preferences
  • the evaluations
  • and eventually the judgment accumulated around all of it

That is my AI knowledge base.

For a company, multiply this across hundreds or thousands of employees and it becomes institutional memory.

This is where I think the idea of "token capital" becomes particularly interesting.

The model should be the most replaceable part

We talk a lot about which company has the best model.

I am increasingly convinced that the model should eventually be the most replaceable part of the stack.

Your knowledge should not be.

I should be able to wake up tomorrow and decide that Claude is better for one problem, ChatGPT for another, Kimi for research, a local model for something sensitive, and some model that does not exist yet for another task.

The model should change.

My accumulated knowledge should follow me.

A knowledge layer I control

What I want is not a vendor retaining every conversation I have ever had.

I want the opposite.

I want a knowledge layer I control.

Something that can understand and preserve what matters from my interactions while letting disposable conversational noise disappear.

A layer containing my projects, decisions, research, workflows, preferences, successful approaches, failures, relationships between ideas and eventually the accumulated expertise created while working with AI.

Then give me permissioning.

  • Let me decide what Claude can see.
  • What ChatGPT can see.
  • What a coding agent can see.
  • What stays completely private.
  • What belongs to a company rather than an employee.
  • What can be forgotten.
  • And what should remain part of the organization's institutional memory for ten years.

The models then become compute sitting underneath my intelligence layer rather than the owners of its state.

The pieces are appearing

We can already see parts of this architecture appearing.

  • NEAR AI is exploring user-owned and private AI infrastructure.
  • Mem0 and OpenMemory are working on persistent memory.
  • Letta is pushing stateful agents and portable memory.
  • OpenRouter makes switching between models easier.
  • MCP is creating a common interface through which different AI systems can reach the same tools and data.

These are important pieces.

But I think there is a larger idea emerging from all of them:

AI needs a sovereignty layer. Not just sovereign models. Sovereign context.

For individuals and companies.

Your model provider should be a dependency you can replace.

Your accumulated intelligence should be an asset you own.

A new kind of vendor lock-in

This becomes even more important inside companies.

A company should not spend five years allowing employees and agents to build enormous amounts of context inside one AI ecosystem, only to discover that changing vendors means rebuilding part of the company's cognitive infrastructure.

That is a new kind of vendor lock-in.

Not lock-in through APIs. Not lock-in through file formats. Lock-in through accumulated understanding.

And that may eventually be much more valuable than either.

The architecture I want

The architecture I want is simple conceptually:

  • My data.
  • My memory.
  • My workflows.
  • My institutional knowledge.
  • My permissions.

Then: whichever model is best for the job.

AI companies can compete aggressively on reasoning, inference, speed, price and capabilities.

Great.

But they should not need to own my memory to win my business.

I rent the intelligence. I should own what I learn with it.