MemFly: On-the-Fly Memory Optimization via Information Bottleneck
A new framework called MemFly, introduced in an arXiv paper (2602.07885v2), addresses the challenge of long-term memory in large language model (LLM) agents. The framework applies information bottleneck principles to enable on-the-fly memory evolution, minimizing compression entropy while maximizing relevance entropy through a gradient-free optimizer. MemFly constructs a stratified memory structure for efficient storage and uses a hybrid retrieval mechanism that integrates semantic, symbolic, and topological pathways with iterative refinement for complex multi-hop queries. Experiments show MemFly outperforms state-of-the-art baselines in memory coherence and response quality. The paper is authored by researchers and was updated as a replacement on arXiv.
Key facts
- MemFly is a framework for on-the-fly memory optimization in LLM agents.
- It is based on information bottleneck principles.
- It minimizes compression entropy and maximizes relevance entropy.
- It uses a gradient-free optimizer.
- It constructs a stratified memory structure.
- It develops a hybrid retrieval mechanism combining semantic, symbolic, and topological pathways.
- The retrieval mechanism includes iterative refinement for multi-hop queries.
- Experiments show MemFly outperforms state-of-the-art baselines in memory coherence and response quality.
- The paper is available on arXiv with ID 2602.07885v2.
Entities
Institutions
- arXiv