The Initiative
We build open specifications, reference implementations with our partners, legal frameworks, and evaluation tools that any organization can adopt.
Active work
An open standard for portable, user-governed personal context that can move across AI systems rather than remain locked inside one platform, with clear permission boundaries and accountability when agents act on a person's behalf.
Loyalty sandboxes, an open agent-rating service, and MemoryArena test whether agents live up to their claims.
Focuses on mitigating systemic risks and emergent collective failures in micro- and macro-scale, multi-agent AI systems.
Legal and technical frameworks translating duties of care and loyalty into concrete agent design and deployment practices.
Human Context Protocol
HCP is an open standard for portable, user-governed personal context: preferences, history, and constraints expressed in a form that can move across AI systems.
It defines structured natural language for everyday preferences, verifiable credentials for higher-stakes categories, and contracts governing search, updates, access, revocation, and audit trails.
Reference partners Turn.io, Learning Economy Foundation, Consumer Reports, and UT Austin test the specification across multilingual, low-connectivity, education, commerce, reputation, and memory-portability use cases.
How context moves between agents
Decide who sees what
Shared: Language, learning
Not shared: Budget, health
Shared: Language, budget
Not shared: Learning, health
Every access is logged. Revoke any agent at any time.
You tell one agent you prefer Spanish and learn best with visuals. With your OK, it saves that to your context.
A new tutoring agent asks for access. You share your language and learning preferences, nothing else.
It starts in Spanish with visual lessons from day one. It can't see your budget or health.
You can see each access in your log and revoke any agent at any time.
Published foundations
AISPA and the System Prompt Index audit the hidden instructions that shape how AI products and agents behave.
A benchmark showing strong recall scores do not guarantee that memory can guide real sequential decisions.
Hardjono, Pentland and Josleyn apply HCP to patient AI conversation memory inside the Saturn digital health ecosystem, a privacy-first architecture built on encrypted citizen data vaults and confidential agent computation.
Maturity path
Build and test specifications against real implementations with partners who surface problems a document alone cannot reveal.
Use lessons from early implementations to shape the next version for a larger, hand-picked group of reviewers.
Circulate specifications publicly after testing and review by multiple independent implementers.
Bring mature work into an established body such as IEEE or W3C for broader consensus and stewardship.