Cognitive State Compilation: Resource-Bounded Construction of Reasoning States for Long-Horizon LLM Agents

Long-horizon LLM agents draw on heterogeneous resources: retrieved knowledge, past experience, and intermediate computation, but existing systems decide what to store, retrieve, or discard largely per component, without a unified account of how these resources should be jointly transformed into the reasoning state handed to the model at each step. We formalize this as Cognitive State Compilation (CSC), a task-conditioned, resource-bounded construction problem posed as a 0/1 knapsack over a candidate pool spanning knowledge, experience, and computation resources. Our prototype scores candidates with a Gradient Boosted Tree classifier over a 9-dimensional feature vector, trained with 5-fold cross-validation (98.6% accuracy, 94.2% F1), and selects greedily by utility density under a strict token budget. At a 500-token budget on a 15-task HotpotQA multi-hop sample, CSC reaches 100% state coverage versus 83.3% for standard RAG and 60.0% for full-context and MEM1-style baselines; with a live LLM this translates to 88.9% F1 versus 66.7% for the strongest baseline. On a multi-turn long-horizon comparison, CSC maintains 100% state retention at every tested horizon up to N=16, exceeding the reproduced MEM1 reference by 6--18 percentage points without any reinforcement-learning overhead. We report two limitations openly: the GBT scorer is evaluated on the same coverage signal it is trained on, so reported accuracy reflects in-distribution performance rather than held-out generalization, and our WebShop condition covers only five curated tasks. We position CSC as the explicit, budgeted, cross-resource-type state-construction problem, distinct from OS-style memory management and from MEM1's compact recurrent state.