Artificial intelligence is rapidly transforming scientific research, bringing us closer to fully autonomous “self-driving” laboratories. However, today’s AI agents face a major bottleneck: amnesia. When running long experimental campaigns or switching tasks, AI models often forget past lessons or get overwhelmed by chaotic conversation logs.
To solve this challenge, our research team at NanoSmartLab (ISASI-CNR) has developed MOSAN (Multi-Objective State-Action Network)—a new cognitive framework that equips AI agents with long-term memory, strategic reasoning, and real-world adaptability.
Our latest study describing this breakthrough has just been published in Machine Learning: Science and Technology (IOP Publishing)!
💡 Key Highlights of the Study:
- 🧠 No More AI Amnesia: MOSAN introduces dynamic memory (Ghost Nodes) that allows the AI to apply lessons learned from optical spectroscopy experiments to completely different domains, such as electrocardiogram (ECG) heart signals, achieving an impressive 99.30% accuracy without starting from scratch.
- ⚡ 78% Cost Reduction: By replacing cluttered chat history with a streamlined Markdown ledger (SiLA), MOSAN prevents cognitive overload and slashes operational LLM costs by up to 78%.
- 🔬 Adapting to Real-World Sensor Aging: When tested on SERS biosensors aged over 6 months—where signal drift normally breaks conventional machine learning models—MOSAN autonomously engineered custom deep-learning architectures to restore classification performance up to 87%.
- 🛡️ 100% Private & Local: MOSAN runs entirely on-premises using open-weight models (such as Mistral and Gemma), ensuring complete data sovereignty for sensitive biomedical research with zero third-party API fees.
🚀 Why It Matters
MOSAN bridges the gap between raw data processing and true scientific reasoning. By enabling AI agents to critique their own steps, retain strategic memory, and adapt to degraded hardware, this work moves autonomous biophysical laboratories one step closer to secure, low-cost, and continuous scientific discovery.
📖 Read the Paper & Explore the Code
- 📄 Full Open-Access Paper (IOP Publishing): https://doi.org/10.1088/2632-2153/ae9fb7
- 💻 Official GitHub Repository: https://github.com/Docdomee/MOSAN
- 🔬 Research Group: NanoSmartLab, Institute of Applied Sciences and Intelligent Systems (ISASI-CNR).
- ✍️ Authors: Domenico Sagnelli, Bryan Guilcapi, Alessia Milano, Amalia D’Avino, Valentina Marchesano, Massimo Rippa, and Lucia Petti.
No responses yet