EnvirTox — NAMs Live-Evo Research Agent

🏗️ Live-Evo Architecture for NAMs

🔑 Key Innovation: What happened ≠ How to use it

Live-Evo decouples experiences (what happened during past NAMs assessments) from meta-guidelines (how to use those experiences for new tasks). Experiences that consistently help are reinforced; misleading ones are down-weighted. When guidelines underperform, the agent reflects and generates new meta-guidelines — this is the self-evolving loop.

Live-Evo Agent Log

🏦 Experience Bank

Weighted repository of past NAMs assessment outcomes. Successful experiences are reinforced; failures are down-weighted. Your expert adjustments feed back into the agentic memory.

📘 Meta-Guideline Bank

Meta-level heuristics that instruct the agent how to use experiences. Updated when ContrastiveEval shows guideline underperformance. These encode your domain expertise as reusable regulatory strategies.

⚖️ ContrastiveEval

Each task runs twice: once with compiled guideline, once without (baseline). The performance gap determines whether experiences are reinforced or down-weighted — this is the core self-improvement mechanism.

💬 Chat with Your NAMs Agent

Agent Log

💡 Try Asking

🚀 How to Deploy This Agent Live

Complete implementation using AG2 framework (from the Live-Evo authors) + ChromaDB for persistent memory + PubMed/ArXiv APIs for literature scanning. Runs entirely on your local machine.

💡 Recommended Stack for EPA Environment

LLM: Claude Sonnet 4.5 via Anthropic API (best cost/performance for extraction)
Framework: AG2 v0.8+ (Apache 2.0, same team as Live-Evo paper)
Vector DB: ChromaDB (local, no cloud — good for EPA data sensitivity)
Literature APIs: PubMed E-utilities, ArXiv API, Semantic Scholar API
Scheduling: cron on workstation or GitHub Actions for daily runs
Storage: Local SQLite + ChromaDB (no external cloud needed)
Human-in-loop: Email digest with Validate/Reject buttons for expert review

ℹ️ About EnvirTox NAMs Agent

EnvirTox is a research platform for New Approach Methodologies (NAMs) in environmental and regulatory toxicology. This agent implements the Live-Evo architecture (Zhang, Wu, Yu, Wu & Wang, 2026; arXiv:2602.02369) for self-evolving agentic memory.

NAMs domains covered: QSAR Modeling, TTC Methodology, In Vitro Assays, Read-Across, Adverse Outcome Pathways (AOPs), IATA Frameworks, PBPK Modeling, and Omics/HTS.

Architecture: The agent maintains two memory banks — an Experience Bank (weighted past assessment outcomes) and a Meta-Guideline Bank (higher-level heuristics for using experiences). ContrastiveEval runs each task with and without guidelines, reinforcing helpful experiences and reflecting on failures to generate improved meta-guidelines.

References:
Zhang, Y., Wu, Y., Yu, Y., Wu, Q., & Wang, H. (2026). Live-Evo: Online Evolution of Agentic Memory from Continuous Feedback. arXiv:2602.02369.
Wei, T. et al. (2025). Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory. arXiv:2511.20857.

NAMs Live-EvoResearch Agent

NAMs Live-Evo
Research Agent