LeanMem: A New Lightweight Memory Framework for LLM Agents
A recent study introduces LeanMem, a streamlined long-term memory framework tailored for LLM-based agents, tackling the shortcomings of current memory systems. Accessible on arXiv (2608.03463), the research suggests that historical dialogue should be categorized differently, taking into account factors like compressibility, temporal dynamics, and fidelity needs. LeanMem discards low-value information while retaining significant segments as compact profile memory, event memory structured by time, or source-grounded records. During updates, only the evolving event memories are refreshed, preventing unnecessary consolidation of stable profiles and fixed records. This strategy aims to minimize token usage and preserve detailed evidence, ultimately enhancing the reliability of long-term interactions. The authors announced this submission to improve memory efficiency in LLM agents, essential for maintaining interactions and utilizing past history.
Key facts
- Paper available on arXiv with ID 2608.03463
- Proposes LeanMem, a lightweight long-term memory framework
- Argues for differentiated handling of dialogue content
- Filters low-value content and stores informative segments in three memory types
- Selectively updates only dynamically evolving event memories
- Aims to reduce token consumption and prevent loss of fine-grained evidence
- Targets LLM-based agents for sustained interactions
- Announcement type: new
Entities
Institutions
- arXiv