The big problem with self-improving agents is that memory can compound mistakes just as easily as it compounds useful lessons.

Salesforce tested 2 memory-based agent methods and found a pretty uncomfortable pattern.

With WebArena’s default task order, ReasoningBank improved performance by 1.5 points.

Shuffle those tasks, and it dropped by 4.5 points instead.

Why? The default order tended to put easier tasks first, so the agent learned cleaner lessons early.

But memory works both ways.

Agents also saved bad lessons, like recommending APIs in an environment where APIs were impossible, then kept pulling those memories back into future tasks.

Results became more unstable in 71% of cases.

Even giving the memory system better task details and environment feedback recovered only 31% of the drop.

So “learning from experience” is only useful if the agent is learning the right thing.

– arxiv. org/abs/2608.18066

Title: "On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification"