SciGroveBeta
Machine Learning

Mem-: Adaptive Memory through Learning When and What to Generate

Xiaoqiang Wang, Chao Wang, Hadi Nekoei, Christopher Pal, Alexandre Lacoste, Spandana Gella, Bang Liu, Perouz Taslakian

Featured May 27, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Instead of just looking up old notes, a new AI memory system called Mem-π learns to smartly create fresh advice for itself, even knowing when to stay quiet if no help is needed.

In depth
The paper introduces Mem-π, a novel framework for adaptive memory in large language model agents that replaces static retrieval with a generative policy. This policy learns to dynamically decide both *when* to produce guidance and *what* specific guidance to generate, using a decision-content decoupled reinforcement learning objective to ensure relevance and prevent unhelpful interventions.

Key Takeaways

  • 1
    Introduces generative memory for LLM agents, moving beyond static retrieval to dynamically create context-specific guidance.
  • 2
    Employs a two-stage distillation process: supervised learning from an offline experience bank followed by reinforcement learning for adaptive generation.
  • 3
    Proposes a decision-content decoupled RL objective that explicitly learns *when* to generate (abstain vs. generate) and *what* content to produce, significantly improving task success and token efficiency.

Conceptual Flow

HIGH LEVEL
1
How Mem-π Learns Adaptive Memory

The system first learns from past examples, then practices making smart choices about when and what advice to give, getting better with feedback.

Past Examples
Agent's Task
Learn & Practice
Smart Advice
Know When to Stay Quiet
2
Mem-π Outperforms Old Memory Methods

This new memory system helps AI agents succeed much more often, especially on tricky tasks, while using less unnecessary information.

Old Memory Methods
Mem-π Memory
Compare Performance
Higher Success Rate
Less Wasted Info