Ensemble QSP: Hierarchical memory defeats context limits for multi-agent AI

Large language models (LLMs) excel at reasoning but struggle with long-horizon research workflows due to their stateless architecture. Ensemble QSP solves this with a three-layer hierarchical memory that keeps injected context bounded and constant—median 301 tokens, max 4,050 across 104 runs—by capping each state category and evicting completed work. The system deploys five specialist worker agents overseen by a domain-expert principal investigator (PI) agent that enforces physical constraints using physics-based checklists and structured domain knowledge. This design allows continuous autonomous operation without context degradation, a critical bottleneck for multi-session scientific tasks.

Benchmarking on pharmacokinetic-pharmacodynamic (PK/PD) model selection shows robust autonomous performance without human intervention, with consistent quality across both cheaper and frontier LLMs. The framework improved PK parameter recovery compared to single-agent baselines and maintained stable model selection across linguistically diverse prompts of the same task. Ablation studies on physiologically based pharmacokinetic (PBPK) models spanning a broad complexity range revealed that PI oversight boosts debugging efficiency while preserving accuracy. The architecture is domain-agnostic—adding a new scientific field only requires a new PI agent configuration—making it a scalable foundation for autonomous research in quantitative biology and beyond.