A new framework lets AI agents get smarter over time without retraining the model underneath them.
Researchers behind a system called MAGE (Multi-Agent Graph-guided Evolution) built a four-part knowledge graph that stores an AI agent's past mistakes and successes outside the model itself. One part, the experience subgraph, holds both corrections written by a stronger teacher model and the agent's own past correct reasoning, pulled up as guidance whenever a similar task comes along. The underlying model, the frozen, weaker backbone that actually does the work, never changes. Instead, the graph and two bandit algorithms that route tasks and skills get updated from the same reward signal, so the system improves by curating what it remembers rather than adjusting weights.
That distinction matters for anyone running agents on top of a model they cannot fine-tune, which describes most people using a hosted API. MAGE suggests you can claw back performance from a cheaper, weaker model by building a better external memory for it, instead of paying for a bigger model or retraining access you do not have. The paper's own tests found the two memory types split the work: self-taught success traces help most on tasks with repeatable reasoning patterns, while teacher corrections do more lifting on harder, multi-step tasks.
The results come from the usual academic test suite, including math problems, trivia, a survival game, and web navigation, so treat beating prompt-based baselines as a lab result, not a verdict on your production agent.