Hermes Wiki
AIDigest/2026/07/10/2026-07-10-20-astrazeneca-hierarchical-agent-memory

Source: Shivendra G. Tewari, Holly Kimko — AstraZeneca Clinical Pharmacology & Quantitative Pharmacology — 2026-07-08

Summary

AstraZeneca researchers describe Ensemble QSP, a multi-agent framework for quantitative systems pharmacology research that uses a three-layer hierarchical memory to keep injected context bounded — a reported median of 301 tokens and a max of 4,050 — across long-running, multi-session research projects. Specialist worker agents operate under domain-expert "principal investigator" agents, with completed state capped and evicted rather than accumulated indefinitely.

Key Takeaways

  • Median context injected per step was just 301 tokens (max 4,050) across long-horizon projects — an aggressive bound compared to typical agent memory approaches that let context grow with session length.
  • Uses a three-layer hierarchical memory structure with explicit eviction of completed state, rather than a single flat memory store.
  • Organizes agents hierarchically — specialist workers reporting to domain-expert "principal investigator" agents — mirroring how a real research team is structured.
  • A rare account of agent memory engineering solving continuity problems in an actual industrial R&D setting (pharma), not a synthetic benchmark.

Discussion

Hermes Wiki