AdaMEM: Test-Time Adaptive Memory for Language Agents

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

The Adaptive Memory Agent (AdaMEM) is a novel framework addressing the rigidity of static memory in language agents during test-time adaptation. It employs a hybrid memory architecture, combining offline long-term trajectory memory with dynamic, on-the-fly short-term strategy memory to guide decision-making. This approach allows agents to continuously adapt to evolving conditions without online model parameter updates, offering a trade-off between token efficiency and adaptability through modes like AdaMEM-low and AdaMEM-high. AdaMEM significantly outperforms static memory baselines, achieving relative gains of up to 13% on ALFWorld and 11% on WebShop, with strong performance on HotpotQA. Further enhanced by Step-MFT, a fine-tuning technique for synthesizing high-quality strategies, performance gains reach up to 17% on ALFWorld and 13% on WebShop.

Key takeaway

For AI Scientists and Machine Learning Engineers developing language agents for dynamic, long-horizon tasks, relying solely on static memory retrieval will limit your agent's adaptability. You should explore AdaMEM's hybrid memory architecture, which dynamically synthesizes context-aware strategies at test-time. Implementing its Step-MFT technique can further train your agent to generate high-utility strategies, potentially yielding up to 17% relative performance gains on benchmarks like ALFWorld and 13% on WebShop, ensuring robust post-deployment self-evolution.

Key insights

Language agents adapt better by dynamically synthesizing context-aware strategies from long-term memory at test-time, rather than relying on static retrieval.

Principles

• Decouple memory storage from strategy abstraction.
• Synthesize strategies dynamically, conditioned on current state.
• Reinforce strategies that actively alter agent actions.

Method

AdaMEM uses offline long-term trajectory memory to dynamically synthesize state-specific short-term strategies. Step-MFT fine-tunes the policy by filtering successful trajectories where the strategy changed the agent's action, using a computation-free proxy for strategy advantage.

In practice

• Employ AdaMEM-low for token-efficient strategy refreshing.
• Use AdaMEM-high for maximum adaptability via step-wise regeneration.
• Apply dual-filter rejection sampling for strategy fine-tuning.

Topics

• Language Agents
• Test-Time Adaptation
• Hybrid Memory
• Strategy Synthesis
• Step-MFT
• Supervised Fine-Tuning

Code references

• yunx-z/AdaMEM

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

Related on AIssential

• AdaMEM: Test-Time Adaptive Memory for Language Agents — AdaMEM enhances language agent adaptability and performance in dynamic, long-horizon tasks through a hybrid, test-time adaptive memory architecture.
• Memp: Exploring Agent Procedural Memory — Procedural memory, dynamically built and updated, significantly enhances LLM agent performance and efficiency across diverse tasks.
• Joint Agent Memory and Exploration Learning via Novelty Signals — JAMEL trains agent memory and exploration together via novelty signals, enabling effective, token-efficient exploration in open-ended environments.
• Memory in Agentic AI — Memory is foundational for autonomous AI agents to learn, adapt, and pursue goals effectively over time.
• AMA: Adaptive Memory via Multi-Agent Collaboration — AMA uses multi-agent collaboration and hierarchical memory to provide adaptive, consistent long-term memory for LLM agents.
• 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 cs.AI updates on arXiv.org.