TIDE framework turns recommender feedback into agent memory
TL;DR
- TIDE is a training-free framework that evolves a content agent's external memory using recommender feedback signals rather than updating model weights.
- The authors report a +7.75 percentage-point gain on their own Memory Evolution Gain metric in offline temporal replay.
- In an online A/B test for e-commerce membership marketing copy, TIDE lifted unique click-through rate by +5.10% and activation rate by +4.79%.
TIDE, a new framework from a group of recommender-systems researchers, claims a +7.75 percentage-point lift on its own 'Memory Evolution Gain' metric by letting a content-generation agent learn from clicks and conversions rather than retraining on them.
The arxiv paper frames the problem as credit assignment under noise: 'impressions, clicks, conversions, and negative feedback from recommendation systems' arrive delayed and tangled with audience, placement and recommender policy. TIDE scores each memory using 'temporal and semantic credit assignment' plus a responsibility-credit step, then reinforces, mutates, crosses over or evicts entries in a capacity-constrained pool. No model weights are touched.
On a real-world online A/B test for e-commerce membership marketing copy, the authors report a +5.10% unique click-through rate and +4.79% activation rate against a no-memory baseline. On a delayed-label benchmark the paper reports the lowest mean absolute error and root mean squared error against the alternatives it compares.
The paper does not name the e-commerce platform the A/B test ran on, and reports the online lifts as percentages without absolute volumes.
Shared on Bluesky by 2 AI experts
Originally reported by arxiv.org
Read the original article →Original headline: How Can Recommendation Feedback Evolve Agent Memory?