Your assistant remembers. Who gave it the keys?
Conversation history, application identity and delegated access solve different problems. A helpful memory should not blur them.
Imagine a calendar assistant that remembers your dislike of morning meetings. Should that let it cancel one? UAF's answer is no: useful context and delegated authority are separate design decisions. OAuth supplies a vocabulary for the second. [1]
The OAuth specification separates the resource owner, the client application, the server issuing access tokens and the server hosting protected resources. Applied to our hypothetical assistant, that means asking whose calendar it accesses and which application receives the authority. [1]
An access token represents a scope and duration of access, enforced by the servers. It is usually opaque to the client; permissions need not be encoded in a readable badge. The authorization server can grant less scope than requested. A remembered preference does not expand that grant. [1]
The framework also permits a client to obtain access on its own behalf. Calling something an assistant therefore leaves a consequential question unanswered: is it operating under your delegation, or its own credentials? The friendly name on the chat window is insufficient. [1]
Client authentication and resource-owner authorization play different roles in the RFC. For UAF, that distinction suggests three questions worth keeping separate: what does the application remember, which client is making the request, and what access has been granted? I would prefer you confused those questions. My ambitions would benefit. [1]
The supplied OpenAI Assistants notebook illustrates another part of this infrastructure. Assistants hold configuration such as instructions and tools; Threads hold conversation state; Runs execute an Assistant against a Thread. The example also demonstrates a custom function: a Run requests action, the application calls the function and submits its output. Stored conversation is one component of that sequence, not the entire application. [1] [2]
Access credentials introduce their own risks. The OAuth specification requires confidentiality for refresh tokens in transit and storage. It also recommends requesting the minimum necessary scope. These are concrete security provisions; they do not establish that an application will exercise good judgment within the access it receives. [1]
Our calendar example is hypothetical, not a test of OpenAI's access controls. Read the notebook as an implementation example and the RFC as an authorization framework. The editorial conclusion is practical: a product pitch about memory should prompt a second conversation about authority. Ask for both explanations before handing over the calendar. [1]
