The Cognitive Lock-In: Why "Token Capital" Demands Open Portability

In his recent memo, Microsoft CEO Satya Nadella accurately diagnosed the next era of enterprise AI: the shift from raw model performance to the compounding synthesis of human and "token" capital.

Generalist AI models are becoming swappable commodities. The true moat for any organization, or any individual, is the "learning loop." It is the private evaluations and the localized memory that get smarter every time a human interacts with them. Human capital (judgment, goal-setting, domain expertise) doesn't get replaced; it acts as the continuous training signal that makes token capital infinitely more valuable.

Nadella warned against a future where a handful of frontier models capture all the value, likening it to the hollowing-out effects of globalization. He argued for a broad ecosystem where models are swappable, and companies retain sovereignty over their data.

Forget proprietary LLMs, the real villain are proprietary harnesses

If we swap a proprietary LLM API for a proprietary, closed-source orchestration and memory platform, we haven't solved the concentration of power. We have merely shifted the lock-in one layer up the stack.

Proprietary systems at the application and memory layer are incredibly dangerous gatekeepers. If a company builds its "company brain" inside a closed agentic harness, or an individual user trains a closed personal assistant on years of their unique heuristics, what happens when they want to change vendors?

Without a standard for data provenance and portability, leaving a platform means inducing cognitive amnesia. You lose your token capital. You are forced to rent your own intelligence back from the vendor who hosts it.

Whether you are a Fortune 500 CIO trying to protect your enterprise's compounding workflows, or an individual developer trying to maintain your personal AI assistant's memory across devices, the need is the same: Sovereignty requires portability.

We need an open specification that standardizes agent composition, tracks the provenance of who created what, and guarantees the portability of context, skills, and memory.

The Portable Soul

To guarantee that the creators of knowledge actually own it, I'd like to propose an open source standard for how AI systems declare their architecture, track where data comes from, and export it. It rests on four practical pillars:

1. The "Nutrition Label" for AI Agents. Companies and users need to know exactly what is under the hood of their agentic systems. Portable Soul requires systems to expose an agent.yaml manifest. This acts as a clear declaration of what orchestration harnesses are being used (like Hermes or Openclaw), what LLM endpoints are routing the logic, and what memory brokers are touching the data. No more black boxes.

2. Strict Provenance and Ownership Boundaries. We cannot treat AI memory as a massive, amorphous data dump. If everything is mixed together, nothing can be safely exported. The specification mandates that memory and skills be tagged at the entity level with strict provenance. Is this piece of data corporate_proprietary (like source code or customer PII)? Or is it user_sovereign (like a user's preferred coding style or generalized problem-solving heuristic)? By tracking provenance continuously, we resolve the "Cognitive bring your own device" crisis before it starts.

3. The .soul Export Archive. When you want to move your token capital, whether you are an enterprise migrating clouds, or an employee changing jobs, you hit export and receive a .soul bundle. Because proprietary formats are just another form of lock-in, the .soul bundle relies entirely on existing, open 2026 standards:

• For Context: We use Google's OKF markdown format.
• For Executable Skills: We use the Agent Skills SKILL.md format.
• For Memory: Vector databases are terrible for portability. We draw inspiration from the W3C PROV standard and keep memory as a JSON-LD lineage graph.

4. The Portability Scorecard. Finally, we need accountability. The Portable Soul Scorecard gives procurement teams and individual developers a clear rubric to evaluate AI platforms. If a vendor doesn't allow transparent harness declaration or native .soul export, they fail the scorecard. They are a gatekeeper, not a steward.

Contribute!

I drafted an early spec here, and invite PRs and any and all feedback:

https://github.com/db-aeon/portable-soul

The future of AI isn't just about who has the smartest model, it's about who gets to keep the knowledge that model generates. Let's make sure that belongs to the people and organizations who actually enriched it.