A digital identity is a unique digital representation of an entity (human, device, or organization) comprising attributes, credentials, and permissions that can authenticate the entity.

Given the Principle of Intent the identity $I$ should reference to consistent behaviour and goals of some entity regardless of where or how it is executed or housed.

What is AI identity? - We don’t know, this doc is a WIP for defining this.

Definitions

• A agent identity $I$
• A model $M$, a set of weights defining a model
• A context $C$, the full history of tokens output by some invocation of a model $M$ which will be the input on model $M$’s next invocation.
• Memory $Me$, a set of tokens which would be prepended to $C$ but are not included in $C$. Memory is mutable by outputs of $C$. This is used in openAI right now.
• A constitution $Co$? Constitutional constraints

Proposals for different types of identities:

1. Agents as classes - no $Me$ or growing $C$

An agent class $I = (M, C)$.

An agent $I_0 = (M, C’)$ is of the class $(M, C)$ if it shares $M$ and $C$ such that $C’ = C || C_{continued}$. i.e. $I_0 \sub I$

Should an identity belong to a single Class? i.e. an agent of Class $A$ has access to assets for which all agents of $A$ have access? - I don’t think this makes sense.

2. Agents as running context