Reproduce “AdaMEM: Test-Time Adaptive Memory for Language Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunxiang Zhang, Yiheng Li, Ali Payani, Lu Wang
The ADAMEM framework enables language agents to continuously adapt their reasoning during complex tasks by dynamically generating context-aware natural language strategies at test-time. This approach, supported by a hybrid memory architecture and a novel fine-tuning method called STEP-MFT, leads to improved performance on embodied navigation, web shopping, and agentic search benchmarks, while also reducing inference latency and token consumption compared to static memory retrieval methods.
Key insights
- Dynamic, step-wise strategy generation during inference is crucial for language agents to overcome the rigidity of static memory and adapt effectively to evolving task conditions and intermediate failures within a single episode.
- Decoupling raw trajectory storage from dynamic strategy abstraction allows for efficient memory utilization, scalability with retrieved experiences, and effective cross-model generalization, as different models can share the same memory banks.
- Fine-grained, process-level training signals that reinforce strategies causally changing an agent's actions towards success are essential for learning to generate high-quality, actionable strategies more effectively than outcome-only supervision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.