Researchers have a fix for a problem plaguing AI agents that learn on the job: teaching them something new keeps wiping out what they already knew.
The issue comes up when an agent's "skill" - a chunk of instructions or memory that guides its behavior - gets rewritten every time it encounters a new task. A paper posted to arXiv on September 30, 2026 describes Semantic-Scope Projected Evolution (SSPE), a method that treats each proposed skill revision the way continual-learning systems treat a gradient update: it checks whether the change would interfere with skills the agent already picked up, then adjusts the update so it fits alongside them instead of overwriting them. The authors borrow the underlying idea from Orthogonal Gradient Descent, a technique used to fight forgetting in neural network training, and adapt it to work on plain natural-language skill text rather than model weights. In tests across synthetic task streams and what the paper calls "heterogeneous real-agent benchmarks," the evolved skills kept performing well on old tasks while picking up new ones, and reportedly held up even when transferred to a different underlying model.
That last point is the more interesting claim. Most agent frameworks today just append or rewrite instructions as they go, which means today's fix can quietly become tomorrow's regression. If a single skill file can absorb updates without a human curating it and still generalize to a different base model, that's a real step toward agents that improve themselves over long stretches of use rather than degrading with every patch.
The paper doesn't report specific accuracy numbers or say which models or benchmarks it used by name, so it's hard to judge how large the gains are or whether they hold up outside the paper's own test setup. Worth watching is whether builders of production agent frameworks actually adopt something like this, since forgetting is exactly the kind of quiet failure that's easy to ignore until an agent starts breaking tasks it used to handle fine.