MemArbiter: A New Framework to Bridge the Memory-Action Gap in LLM Agents
A new framework called MemArbiter has been developed by researchers to enhance the decision-making capabilities of large language model (LLM) agents during long-horizon tasks. This framework tackles the 'Memory-Action Gap,' a phenomenon where relevant information remains unutilized in decision-making due to inadequate formation, organization, prioritization, or presentation. MemArbiter breaks down interaction histories into atomic components and categorizes them into five functional Memory Banks, while dynamically adjusting memory salience based on bank demand, item relevance, focal-ambient representations, and a temporal presentation gate. Evaluated on the ALFWorld benchmark against Flat Retrieval and Flat Recency methods with consistent per-step memory budgets, MemArbiter demonstrated improved performance using an open-weight action-generation model, as detailed in the paper available on arXiv (identifier 2608.02113).
Key facts
- MemArbiter is a function-aware memory arbitration framework for LLM agents.
- It addresses the Memory-Action Gap, a post-access failure in long-horizon tasks.
- The framework decomposes interaction histories into atomic items.
- It organizes items into five functional Memory Banks.
- It uses bank-level demand, item-level relevance, focal-ambient representations, and a temporal presentation gate.
- Evaluation was conducted on ALFWorld against Flat Retrieval and Flat Recency.
- The study uses an open-weight action-generation model.
- The paper is available on arXiv with ID 2608.02113.
Entities
Institutions
- arXiv