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

2026-05-20 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

Mem-$π$ is a novel framework designed to provide adaptive memory for large language model (LLM) agents by generating context-specific guidance on demand. Unlike traditional memory-augmented agents that rely on similarity-based retrieval of static entries, Mem-$π$ employs a dedicated language or vision-language model, separate from the main agent, to dynamically decide both when and what guidance to produce. This system is trained using a decision-content decoupled reinforcement learning (RL) objective, allowing it to abstain from generating guidance when it would not be beneficial and otherwise produce concise, useful information. Across diverse agentic benchmarks, including web navigation, terminal-based tool use, and text-based embodied interaction, Mem-$π$ consistently outperforms existing retrieval-based and RL-optimized memory baselines, demonstrating over 30% relative improvement on web navigation tasks.

Key takeaway

For machine learning engineers designing LLM agents that require adaptive memory, you should explore dynamic guidance generation frameworks like Mem-$π$. This approach, which learns when and what to generate, offers significant performance gains, particularly over 30% in web navigation, compared to traditional similarity-based retrieval. Integrating a dedicated guidance model can enhance agent autonomy and task performance by providing context-specific, on-demand support.

Key insights

Mem-$π$ adaptively generates context-specific guidance for LLM agents on demand, outperforming static retrieval methods.

Principles

• Guidance generation should be context-specific.
• Decouple decision to generate from content.
• Abstain from guidance when unhelpful.

Method

Mem-$π$ employs a dedicated language or vision-language model, separate from the agent, to jointly decide when and what guidance to generate. It is trained with a decision-content decoupled reinforcement learning objective.

In practice

• Improve web navigation task performance.
• Enhance terminal-based tool use.
• Boost text-based embodied interaction.

Topics

• LLM Agents
• Adaptive Memory
• Reinforcement Learning
• Guidance Generation
• Web Navigation
• Tool Use

Code references

• WujiangXu/A-mem
• WujiangXu/A-mem-sys

Best for: Research Scientist, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Mem-$π$: Adaptive Memory through Learning When and What to Generate — Mem-$π$ enables LLM agents to generate context-specific guidance adaptively, outperforming static retrieval methods.
• SWE-MeM: Learning Adaptive Memory Management for Long-Horizon Coding Agents — Long-horizon coding agents can achieve superior performance and efficiency through adaptive, proactive memory management.
• Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents — HORMA uses a hierarchical, file-system-like memory to enable efficient, detailed context retrieval for LLM agents in long-horizon tasks.
• MemRefine: LLM-Guided Compression for Long-Term Agent Memory — LLM-guided compression, MemRefine, manages agent memory by using an LLM judge for factual content-based decisions, outperforming surface similarity.
• GUIDE: Guided Updates for In-context Decision Evolution in LLM-Driven Spacecraft Operations — GUIDE enables LLMs to adapt in real-time by evolving natural-language decision rules without weight updates.
• A 0.12% parameter add-on gives AI agents the working memory RAG can't — Delta-mem offers an efficient, dynamically updated associative memory for AI agents, overcoming RAG and context window limitations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.