Mem-π: Adaptive Memory for LLM Agents Generates Context-Specific Guidance
Mem-π is an innovative system designed to improve large language model (LLM) agents by introducing flexible memory that provides timely, relevant help instead of just using fixed information from outside sources. Unlike traditional memory-augmented agents that often pull from past experiences or skill sets, Mem-π uses a unique language or vision-language model with its own parameters. This model evaluates the agent's current situation to decide the most helpful advice to give. It’s been trained with a special reinforcement learning approach that helps it avoid giving irrelevant responses, ensuring it offers clear and useful guidance. The effectiveness of this framework has been tested across several benchmarks, including web navigation and terminal tasks.
Key facts
- Mem-π is a framework for adaptive memory in LLM agents.
- It generates guidance on demand rather than retrieving from external memory.
- Existing methods rely on similarity-based retrieval that often misaligns with context.
- Mem-π uses a dedicated language or vision-language model separate from the downstream agent.
- The model jointly decides when and what guidance to produce.
- It is trained with a decision-content decoupled reinforcement learning objective.
- The framework can abstain from generating unhelpful guidance.
- Tested on benchmarks including web navigation and terminal tasks.
Entities
Institutions
- arXiv