SLICE-Mem cuts LLM memory costs by 99.5% in new framework
Researchers introduced SLICE-Mem, a lightweight memory framework for LLM agents that reduces reliance on generative models for memory tasks. The system uses salience-guided admission and predicate-centered extraction to filter and organize user interactions without real-time LLM calls. Instead, it defers consolidation to an offline stage, cutting token consumption by 92.4% on Memora Weekly and…
Key points
- SLICE-Mem uses salience-guided admission and predicate extraction to filter user turns without real-time LLM calls
- Reduces token consumption by **92.4%** on Memora Weekly and **99.5%** on LongMemEval-S vs. MemOS baseline
- Achieved **46.60** FAMA on Memora Weekly and **56.8%** accuracy on LongMemEval-S with Qwen backbones
Evaluated on subsets of Memora and LongMemEval-S, SLICE-Mem achieved 46.60 FAMA on Memora Weekly and 56.8% accuracy on LongMemEval-S using 8B and 30B Qwen backbones. While slightly trailing MemOS’s 46.68 and 59.3%, the framework demonstrates efficiency by minimizing LLM inference overhead while maintaining coherence in long-term agent interactions.
The story so far
3 episodes →- SLICE-Mem cuts LLM memory costs by 99.5% in new frameworkthis story
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers propose RSD metric to measure LLM agent consistency · 1 src
- Researchers propose tracequery to cut diffusion LLM inference costs · 1 src
- Researchers propose knowledge-guided LLM for 6G service orchestration · 1 src
- Arxiv paper introduces Boa Compressor for Linear-Time context compression · 1 src
- Embedding-based subset selection improves ASR Fine-Tuning on hebrew and danish · 1 src
Comments
via GitHub Discussions