Long‑horizon personalized agents must continuously decide what to add, modify, or discard in their memory. Learning an optimal update policy is difficult because conventional training relies solely on terminal question‑answering accuracy, providing a coarse feedback signal that treats all intermediate updates equally. Moreover, the utility of an update is inherently user‑specific—identical memory actions can have opposite effects across different user personas. Existing approaches use a globally shared evaluation standard with at most a single scalar intermediate reward, failing to capture this heterogeneity.
To tackle these issues, we introduce Personalized Memory Rubrics (PMR). PMR translates terminal feedback into fine‑grained, per‑step, per‑user credit assignment. The evaluation is decomposed into two layers:
- Universal layer that captures cross‑persona failure patterns, forming a global baseline.
- Per‑user adaptive layer that models each user’s individual traits as a delta over the universal baseline.
Through Rubric‑Guided Credit Assignment, PMR redistributes credit at the token level: validated decisions on successful trajectories are amplified, while failed trajectories fall back to standard policy optimization to avoid unreliable attribution. The rubric and the memory policy co‑evolve in a closed loop with anti‑circularity guarantees, ensuring stable learning. Crucially, PMR incurs zero inference overhead, operating without any rubric or external judge during deployment.
Extensive experiments on four benchmarks show consistent gains over strong memory‑augmented baselines, confirming that fine‑grained, user‑aware credit assignment substantially improves long‑term memory learning.
Review: PMR’s two‑layer decomposition provides precise control over memory updates while respecting user differences, offering a promising direction for broader personalized systems.