AI Context

AI agents are becoming capable of producing software, documents, plans, decisions and transactions at negligible marginal cost.

Applications may become temporary. An agent may generate an interface or workflow for a particular task, use it briefly and discard it. But the person or organisation behind that work must remain.

Something durable still needs to hold:

• identity;
• knowledge and history;
• relationships and commitments;
• preferences and intentions;
• authority delegated to agents;
• evidence of what was done;
• responsibility when something goes wrong.

Today, most of that continuity is being accumulated inside proprietary AI services. The agent that has known you longest becomes difficult to replace, even when a better agent appears. This is application lock-in reproduced at a deeper level.

This project explores a different model:

Identity → context and authority → replaceable agents

The person or organisation is the durable root. Agents are temporary clients acting with limited, inspectable authority.

Why this matters

The economic value of producing competent digital work may decline as AI makes that work abundant. What remains scarce is not necessarily production itself, but:

• trust;
• legitimate authority;
• accountability;
• institutional memory;
• continuity over time.

The important question will increasingly be less:

Who made this?

and more:

Who authorised it, under what rules, using which information, and who is accountable for the result?

Human oversight cannot mean approving every action manually. Instead, people and organisations need to define the purpose, boundaries and escalation rules within which agents may operate.

Humans provide the constitution, not every keystroke.

What this project proposes

The project began as an investigation into portable AI memory: how a person could export the context accumulated by one AI service and introduce it to another without starting again.

That remains necessary, but it is not sufficient.

A complete principal-controlled agent environment needs several distinct layers:

Context

The durable knowledge an agent may use: source material, conversations, projects, preferences, decisions and relationships.

Agent-created summaries and inferences must remain distinguishable from authoritative source material. They must be inspectable, correctable and replaceable.

Authority

A machine-readable mandate describing what a particular agent may do on behalf of a person or organisation.

Authority should be explicit, limited, revocable and independent of the agent vendor.

Accountability

A record of which agent acted, under which mandate, using which information, with which approvals, and what resulted.

Review

Agents will inevitably form incomplete or incorrect interpretations. The REVIEW.md roundtrip makes those interpretations visible so the principal can approve, reject or replace them.

This is not only a portability feature. It is memory hygiene and governance.

Reuse the Web

The project follows the Web’s usual design method: reuse as much as possible and add only the minimum missing glue.

The intended architecture builds on existing standards and protocols:

• WebID and Solid for durable identity and principal-controlled storage;
• linked data for context that can cross application boundaries;
• Solid Application Interoperability for data-access needs and grants;
• ODRL for permissions, prohibitions, constraints and duties;
• OAuth and AuthZEN for runtime authorization;
• MCP and A2A for agent interaction;
• PROV-O for execution provenance;
• Verifiable Credentials and Data Integrity where portable cryptographic evidence is required.

The main proposed addition is a small ODRL Agent Mandate Profile: a constrained way to express authority delegated by a principal to a specific software agent.

The mandate is durable. Runtime credentials are temporary projections of it. Agents can be replaced without transferring ownership of the principal’s identity, context, authority or history.

Design principles

1. The principal is the durable root. A person or organisation owns its identity, context, mandates and evidence.

2. Agents are replaceable. No agent vendor should own the continuity of the principal it serves.

3. Context and authority are separate. Permission to know something does not imply permission to act. Permission to act does not imply unrestricted access to context.

4. Inferences are not facts. Agent-created interpretations retain their provenance and can be reviewed or corrected.

5. Authority is explicit and revocable. Agents receive bounded mandates rather than general permission to impersonate the principal.

6. Unknown means deny. An evaluator that cannot understand or verify a rule must not infer permission.

7. Important actions leave evidence. It should be possible to reconstruct what was authorised, what happened and who remains accountable.

8. The standards remain open. The goal is interoperability, not a new platform that owns the same relationship.

Repository

• docs/appocalypse.md describes how AI may remove the scarcity on which the software application industry is based.
• docs/design.md describes the user-owned AI context model and its linked-data representation.
• docs/agent-mandate.md proposes an ODRL profile for bounded, revocable agent authority.

The project is exploratory. Its purpose is to discover the smallest set of open conventions and tools needed for people and organisations to remain in control while increasingly capable agents perform most digital work.

No replaceable intelligence should own the continuity of the person or organisation it serves.