MAGE: Multi-Agent Self-Evolution with Co-Evolutionary Knowledge Graphs

MAGE introduces a novel framework for self-evolving language model agents that uses co-evolutionary knowledge graphs to preserve learned knowledge across iterations without modifying the base model. The system externalizes learning into structured memory subgraphs, enabling frozen backbone models to improve through retrieved guidance while maintaining inference stability across nine diverse benchmarks.

MAGE addresses a fundamental challenge in AI agent development: how to enable continuous learning without destabilizing frozen model architectures. Traditional approaches rely on natural language feedback or episodic memory, which create inference-time complications and knowledge retention problems. This research proposes externalized knowledge graphs as a solution, separating the learning mechanism from the execution model itself.

The framework's architecture reflects broader trends in AI systems design toward modular, interpretable learning. By maintaining separate subgraphs for experience storage, task-level routing, and skill-level decisions, MAGE creates a system where learning dynamics remain observable and controllable. The append-only memory design prevents catastrophic forgetting while bounded curriculum coverage ensures stable improvement—addressing core challenges that have plagued continual learning systems.

The empirical validation across nine diverse benchmarks demonstrates practical utility beyond narrow domains. Performance gains span mathematical reasoning, question answering, financial analysis, and even open-world gaming environments, suggesting the approach generalizes meaningfully. The ablation studies showing complementary contributions from self-harvested traces versus teacher corrections indicate sophisticated knowledge composition mechanisms.

For AI systems practitioners, MAGE's approach has implications for deployment scenarios requiring frozen models in production while supporting background evolution. The structured knowledge graph representation enables auditing learned behaviors and understanding failure modes—critical for safety-critical applications. The framework's stability properties make it relevant for systems where inference-time unpredictability creates unacceptable risks.

• →MAGE enables frozen-backbone language models to improve through self-evolution without modifying core weights
• →Co-evolutionary knowledge graphs separate learning mechanisms from execution, improving interpretability and stability
• →Self-harvested success traces and teacher corrections provide complementary learning signals across different task types
• →The framework demonstrates consistent improvements across nine diverse benchmarks spanning reasoning and navigation domains
• →Append-only memory architecture prevents catastrophic forgetting while maintaining inference-time determinism