Harness Continual Learning Reframes Agent Memory

The important shift is treating the agent harness as a learnable system—and its edits as a new source of catastrophic forgetting.

• –Separates candidate generation from commitment, enabling updates to be tested before deployment
• –Tracks retention alongside improvement instead of blindly accumulating memory
• –Gives developers a practical path to self-improving agents without model retraining
• –Makes prompts, skills, routing, and memory first-class engineering surfaces
• –The main open question is whether these gains survive messy, adversarial production workloads