Frontier models are becoming Turing machines for intelligence: general-purpose systems that can apply knowledge supplied at runtime. The same model can reason about your software architecture or write in your company's voice. What changes is the knowledge and intent you supply.
This is in-context learning. The model uses explanations and examples in its context without changing its weights. A decision you made five minutes ago can guide its next action, even though that decision could never have appeared in its training data.
A general-purpose computer becomes useful through the software it can run. As models become more capable of applying what we teach them, we need a way to make that teaching reusable.
Teaching a subject means teaching its prerequisites
Let's say you want an agent to build an invitation system. It needs to understand membership, which depends on your definitions of Account and identity. At Socra, a person's identity exists independently of their email address and the Accounts they belong to. An agent could build a perfectly conventional invitation system around different assumptions unless we supply ours.
Giving it the invitation rules alone leaves those assumptions unresolved. Copying every prerequisite into every prompt gives us definitions to maintain in multiple places.
Cortex lets us put each subject in a Module and declare the other Modules needed to understand it. These relationships form a Semantic Directed Acyclic Graph, or SemDAG. In a small invitation example, the prerequisite order is:
identity → account → membership → invitations
An agent asks Cortex to Flash invitations. Cortex follows every declared dependency, including transitive dependencies, and assembles the complete closure with prerequisites first. Each Module appears once, even when several paths reach it. Knowledge outside that closure stays out.
Keep what you teach
Now suppose a code review reveals a missing rule about identity. Your agent can preserve the correction and its reason in the identity Module. The next Flash of invitations brings in that correction through the existing dependencies. Another agent working on Account administration can retrieve the same revised definition.
You can explain the correction while working in ChatGPT and have an agent use it later in Claude Code or Codex. Cortex's guidance helps connected agents capture this knowledge as work reveals it, under your instructions and permissions.
Cortex preserves the knowledge outside the conversation, so later work can load its current version.
That also separates our knowledge from the model consuming it. When a more capable model becomes available, it can load the definitions and reasons we have already developed. Our team can continue refining its knowledge while the models improve at applying it.