Current AI coding agents build extensive knowledge to make code changes, analogous to developers building mental models. However, this knowledge is largely lost after each session or commit, similar to how a developer's detailed model disappears after editing.

This "knowledge chipper" effect leads to massive inefficiency. If one LLM, like Claude, builds context and then another, like Codex, needs to work on the same code, it starts from scratch, burning millions of tokens. This loss of institutional knowledge and session resumability is a significant problem, especially for larger teams where code reviewers struggle to grasp the nuance behind changes without the AI's original context.

The urgency for "LLM portability" is highlighted by events like the AWS datacenter disruption in Bahrain, where war made regional AI usage impossible, forcing a rethink of where and how AI context is stored. The author advocates for a system where the rich context built by AI agents is preserved and transferable, preventing the repeated burning of tokens and enabling true continuity of work.