The paper introduces TIDE, a training-free framework that evolves an agent's external memory using delayed, noisy recommendation feedback. It tackles the problem of attributing outcomes (clicks, conversions) to specific memories when feedback is delayed and confounded by audience and placement. TIDE uses three stages—temporal, semantic, and responsibility—to measure memory fitness and then applies reinforcement, crossover, mutation, or eviction to update the memory population under a capacity constraint.
On an e-commerce content-generation agent, TIDE achieved a +7.75-percentage-point improvement in Memory Evolution Gain (MEG) over a no-memory baseline in offline tests, and significantly improved unique click-through rate and activation rate in an online A/B test. It also outperformed baselines on a delayed-label benchmark with the lowest error rates and highest MEG. This work is interesting because it provides a principled method for converting real-world feedback into better agent behavior without retraining the model, which is critical for adaptive systems operating in changing environments.