Proposed architecture for AI code editors: local compression layer tracks programmer intent as vectors sent to remote model for process-aware suggestions.

Current AI coding assistants (like Cursor) excel at state—indexing the codebase and retrieving relevant code—but fail when the right suggestion depends on intent: what the programmer has been trying to do over the last twenty minutes, not what the code looks like now.

The proposed fix uses a small local model running in the editor that watches every edit, maintains a running vector representation of programmer intent, and sends that compressed semantic trace upstream. The remote model can also dispatch investigation tasks back to the local model—for example, verifying a refactoring won't break unseen callers or checking whether conditions for a known bug pattern exist—before generating suggestions.

This enables the tool to offer next-hour guidance instead of next-line prediction: noticing race conditions before they're written, recognizing that a library could eliminate a pattern being manually implemented, or spotting that a new abstraction will collide with a decision made in another file weeks ago.

The embedding space alignment between local and remote models remains unsolved and requires annotated edit sequences with ground-truth intent labels for training, data that does not yet exist at scale.