The Big Picture
Jointly training the parts of a personalized memory system so they share credit for overall answers leads to better personalized responses than optimizing each part separately.
ON THIS PAGE
The Evidence
Treat the memory system as a team orchestrator–worker pattern: a fine-grained extractor, a coarse profile builder, and a retriever were optimized together by modeling their interaction as a sequential process and sharing global reward based on contribution. That adaptive credit assignment aligns local goals (what each agent does) with the global goal (answer accuracy), reducing redundant or missing memory content. On a long-context personalization benchmark, the jointly trained system consistently beat independently trained agents across short and very long histories. Model Context Protocol (MCP) Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1Benchmark scale: PersonaMem contains ~180 long user histories and ≈6,000 multiple-choice personalized queries used for evaluation.
2History sizes tested: experiments run on three conversation-length settings — 32K, 128K, and 1M tokens — covering short to very long user histories.
3Agent design and reward balance: the system uses three specialized agents (extraction, profile, retrieval) and balances construction vs. precision with construction weight α=0.8 and retrieval precision weight β=0.2.
What This Means
Engineers building personalized chat assistants and product leads deciding how to structure memory for long user histories will benefit—this shows a practical way to get multiple memory components to cooperate. Researchers working on multi-agent systems or evaluation should consider joint optimization and adaptive credit as a path to improve end-to-end performance.
Key Figures
Fig 1: Figure 1 : Construction and retrieval agents, optimized on local tasks independently, yield lower global system performance than those under joint optimization.
Fig 2: Figure 2 : Illustration of challenges for joint optimization.
Fig 3: Figure 3 : The overview of our proposed framework, CoMAM, which regularizes agents’ execution as MDP trajectories for joint RL optimization and fosters collaboration via adaptive credit assignment to achieve local-global alignment.
Fig 4: Figure 4 : Detailed performance of different methods across seven question types on the PersonaMem benchmark across three history length settings ( i.e., 32K, 128K, and 1M). The number of question types is consistent with those listed in Section 5.1
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Results are demonstrated on a single, sizable benchmark (PersonaMem); behavior on other datasets and in noisy, real-world logs needs validation. The approach relies on rule-based and frozen-model rewards to measure local and global quality, which may bias learning if those rewards are imperfect. Joint training increases algorithmic complexity and computational cost compared with training components independently; sensitivity to the credit-assignment weight was observed and must be tuned for new deployments. Context Drift
Methodology & More
Memory systems for personalized chat break a user’s long history into stored pieces (fine facts, coarse preferences) and later retrieve them to answer new queries. Optimizing each component by itself can produce conflicts: an extractor may save too much noisy detail or a retriever may over-filter useful information, hurting the final answer. The proposed method models the pipeline as a sequential decision process so the extractor, profile builder, and retriever are treated as steps in one flow. Each agent gets its task-specific reward plus a share of the global answer accuracy, where the global share is assigned adaptively by measuring how well an agent’s local ranking of outcomes matches group-level global results (ranking consistency). Blackboard Pattern Tree of Thoughts Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Multiple authors but no affiliations, h-indexs, or venue prestige provided; looks like an emerging/limited-info work on arXiv.