A Hierarchical Memory Architecture Overcomes Context Limits in Long-Horizon Multi-Agent Computational Modeling

Jul 16, 2026

Large language models (LLMs) excel at reasoning, but their stateless design restricts deployment for long-horizon, multi-session scientific workflows that require continuity and quantitative rigor. Ensemble QSP introduces a three-layer hierarchical memory that keeps injected context bounded and constant over project duration by capping each memory category and evicting completed work. Five specialist agents under principal investigators enforce physical constraints via checklists and structured domain knowledge, enabling continuous autonomous operation. Benchmarks show robust autonomous QSP model selection with improved PK parameter recovery and stable quality across lower-cost and frontier LLMs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

Introduction

The context window bottleneck

Hierarchical memory architecture overview

Results

System architecture

A hierarchical memory architecture overcomes context limits in long horizon multi-agent computational modeling Shivendra G. Tewari and Holly Kimko Systems Medicine, Clinical Pharmacology & Safety Sciences, AstraZeneca, Gaithersburg, Maryland Abstract Large language models (LLMs) demonstrate remarkable reasoning capabilities, yet their stateless architecture fundamentally limits deployment in long-horizon research workflows requiring multi-session continuity and quantitative rigor. Here we present Ensemble QSP, a multi-agent framework featuring a three-layer hierarchical memory architecture that keeps injected context bounded and constant in project duration (mid-term project state: median 301 tokens, max 4,050, across 104 runs) by capping each state category and evicting completed work, enabling continuous autonomous operation without context degradation. The system orchestrates five specialist worker agents under domain-expert principal investigators, enforcing physical constraints through physics-based checklists and structured-domain knowledge. Comprehensive benchmarking demonstrates robust autonomous pharmacokinetic-pharmacodynamic modelselection without human intervention, consistent result quality across both lower-cost and frontier LLMs, improved PK parameter recovery relative to single-agent baselines, and stable model selection across linguistically diverse prompts ofthe same task. Feature-level ablation across physiologically based pharmacokinetic (PBPK) models spanning a broad complexity range shows that PI-agent oversight improves debugging efficiency while preserving final accuracy across conditions. The architecture is structurally domain-agnostic, adding a new scientific domain requires only a new PI agent configuration. Introduction The context window bottleneck Large language models (LLMs) have advanced from text generation tools to scientific reasoning engines capable of autonomous chemical synthesis [1], protein structure prediction [2], and endto-end scientific discovery and its publication [3] . These successes demonstrate sufficient parametric knowledge and reasoning capacity for complex scientific tasks. However, a critical gap persists, real scientific projects are not single-prompt endeavors. They unfold over weeks or months, require multi-session continuity, accumulate intermediate results, and demand adherence to domain-specific physical constraints throughout. Quantitative systems pharmacology (QSP) exemplifies this challenge [4] . A typical QSP modeling workflow involves formulating systems of stiff ordinary differential equations (ODEs), fitting parameters to heterogeneous non-clinical and clinical data from multiple studies, validating against held-out observations, and iterating across drug candidates at the same time maintaining physiological plausibility, unit consistency, and regulatory-grade documentation. No existing LLM agent framework supports such workflows autonomously, because they were designed for task-level completion rather than project-level continuity. Multi-agent system (MAS) frameworks such as AutoGen [5], CrewAI, and LangChainLangGraph enable role-based agent collaboration but treat context management naively: agents either pass full conversation histories (leading to O(N×T) token growth where N is session count and T is tokens per session) or start fresh on each interaction (losing accumulated scientific knowledge) . The AI Scientist [3] demonstrated end-to-end drug discovery followed by its paper generation but operates within a single session on well-scoped machine learning tasks, without addressing the stiff-ODE solving, multi-source data integration, and physical constraint enforcement that characterize mathematical biology. We address this limitation with a hierarchical memory architecture comprising three bounded layers. The short-term layer holds the live working context injected each turn: a fixed window of recent conversation turns (4–20, se

FAQs

What problem does the hierarchical memory architecture solve in long-horizon multi-agent modeling?

It addresses the LLM context-window bottleneck by preventing injected context from growing with the number of sessions. The approach keeps mid-term project state bounded and evicts completed work so continuity is preserved without context degradation.

How does Ensemble QSP structure memory across different time horizons?

It uses three layers: a short-term layer with recent turns plus capped scratchpads and a rolling buffer of recent cross-agent outcomes, a mid-term layer with bounded task-relevant slices in structured JSON plus an auto-summarizing decision log, and a long-term layer containing domain-independent or domain knowledge resources like a QSP handbook and physics checklists.

What evidence supports that the system works autonomously and reliably?

Comprehensive benchmarking demonstrates robust autonomous pharmacokinetic-pharmacodynamic model selection without human intervention. It also reports consistent result quality across different LLM cost tiers and improved PK parameter recovery versus single-agent baselines, with stable model selection across linguistically diverse prompts.

Jul 16, 2026

Upload by Evangeline