CASCADE equips LLM agents with an episodic memory that accumulates, selects, and refines task-relevant examples during deployment—without changing model weights. The system treats experience reuse as a contextual bandit problem, balancing exploration (trying new strategies) and exploitation (using known good examples) with formal no-regret guarantees. Testing across 16 diverse tasks (medical diagnosis, legal analysis, code generation, tool use, embodied interaction), CASCADE improves macro-averaged success by 20.9% over zero-shot prompting and beats both gradient-based fine-tuning and other memory baselines.
Main takeaways:
- CASCADE enables "deployment-time learning" (DTL)—the model improves from experience after training ends, without parameter updates.
- It builds an explicit episodic memory of past cases and uses contextual bandit algorithms to decide which cases to reuse.
- Formal guarantees ensure long-term regret bounds (the system won't get stuck in bad strategies).
- Beats zero-shot by 20.9% on average and outperforms gradient-based and simpler memory baselines across 16 tasks.
- Positions deployment as an adaptive learning phase, not just static inference.