Plan-Anchored Context Compilation: Treating the Prompt as a Compiled Artifact, Not a Container

Expanding the context window has not solved long-horizon planning for LLM agents, because the problem is not storage capacity alone but what is stored, in what form, and when it is retrieved. We analyze why current approaches—full context, rolling summaries, semantic RAG, and hierarchical textual memory—fail to preserve commitments and assumptions across hundreds of steps. We then propose Plan-Anchored Context Compilation (PACC): a typed state outside the LLM, modified by the LLM via state diffs validated by a deterministic validator, with assumption-driven replanning, where the prompt at each step is built as a compiled output from the frontier of the current plan. This is a position paper: we present the argument, the architecture, and an evaluation protocol, without claiming experimental results.