AdaMem: Learning What to Remember for Personalized Long-Horizon LLM Agents

Researchers introduce AdaMem, an adaptive memory system for LLM agents that learns what information to retain based on individual user preferences rather than storing everything. The method achieves up to 9% QA accuracy improvement while reducing memory bloat, addressing practical constraints of inference costs and finite context windows in production systems.

AdaMem tackles a fundamental challenge in deploying long-horizon LLM agents: managing memory efficiently at scale. Traditional approaches attempt comprehensive retention, but production systems face hard constraints from inference costs and context limitations. This research identifies that memory value is inherently personalized—information irrelevant to one user wastes precious context space that could benefit another. The solution uses role-specific memory policies refined through weekly QA feedback and self-reflection mechanisms with rollback capabilities, enabling the system to learn user-specific retention patterns over time.

The problem of memory bloat in dialogue systems has grown acute as applications demand longer interaction histories. Researchers validated their approach through AdaMem-Bench, a benchmark simulating realistic multi-week interactions with incremental feedback. Results show consistent improvements across different extraction models and feedback mechanisms, with memory volume reduction suggesting the system successfully filters noise.

For AI developers building production agents, this research offers practical validation that learned, selective memory outperforms uniform retention policies. The lightweight implementation approach—using self-reflection patches rather than expensive retraining—makes the method accessible for deployment. The framework's emphasis on role-dependency acknowledges that different user contexts require different information hierarchies, a principle applicable beyond dialogue to any agent requiring personalized state management.

Future directions include extending adaptive memory to multi-user scenarios and exploring dynamic policy updates during deployment rather than weekly intervals. The work establishes important baselines for memory management efficiency, potentially influencing how LLM applications prioritize information storage.

• →AdaMem learns user-specific memory policies through feedback, improving QA accuracy by up to 9% versus uniform retention approaches
• →Adaptive memory selection reduces total memory volume by 9% while maintaining or improving retrieval performance
• →Role-dependent memory policies address production constraints of inference costs and finite context windows
• →Lightweight self-reflection with rollback enables efficient policy refinement without expensive model retraining
• →The framework demonstrates that what to remember depends on individual user context rather than universal relevance