Learning to Continually Learn via Meta-learning Agentic Memory Designs

arxiv:2602.07755

Original source · Paper JSON

Agents, open-endedness, AGI · 2026-02-08

Cites (6) · Cited by (4)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5

Contents

• Learning to Continually Learn via Meta-learning Agentic Memory Designs
• Abstract
• Keywords:
• 1 Introduction
• 2 Related Work
• Memory for Agentic Systems.
• AI-generating algorithms and the “Learning to learn” paradigm.
• Automated Design of Agentic Systems.
• 3 Learning of Memory Designs
• 3.1 Search Space for Memory Designs
• 3.2 Evaluation of Memory Designs
• 3.3 Open-Ended Exploration
• 4 Experiments
• 4.1 Experiment Setup
• 4.2 Benchmarks
• 4.3 Baselines
• 4.4 Results
• 5 Conclusion, Safety, and Future Work
• Impact Statement
• Acknowledgments
• Appendix A Algorithmic Details
• A.1 Search Space
• A.2 Evaluation of Memory design
• Memory Collection Phase.
• Deployment Phase.
• A.3 Evaluation Cost
• End-to-End Memory Cost.
• Token Size of Retrieved Knowledge.
• A.4 Memory Design Archive and Sampling Process
• Memory Design Archive.
• Sampling Process.
• A.5 Prompts of Meta Agent
• A.6 Pseudocode
• Appendix B Experiment Details
• B.1 Benchmark Details
• B.2 Baseline Details
• B.3 Additional Details of Learning Process
• B.4 Additional Details of Testing
• Appendix C Result Details
• C.1 Learning Processes on Other Benchmarks
• C.2 Results of Greedy Exploration
• C.3 Example Results of Learned Memory Designs
• Instructions for reporting errors