Long-term memory helps LLM agents support users in a personal and temporally consistent way. But ingesting, extracting and organizing memories often requires repeated model calls, which increases costs and waiting times. According to a paper on arXiv (category cs.CL), SLICE-Mem addresses exactly this problem.
What is SLICE-Mem?
SLICE-Mem is a lightweight memory framework for LLM agents. It combines two ideas: an ingestion step guided by the salience (significance) of an utterance, and an extraction step built around the predicate, that is, the core of a statement.
How does SLICE-Mem work?
A compact classifier decides which user messages are worth storing. An extractor based on semantic dependencies turns these messages into concise memory entries. This online pipeline needs no generative LLM calls.
The reconciliation between multiple entries, for example to resolve overlaps, is moved to a later offline consolidation. This way, the costly work does not occur during the ongoing conversation.
What do the tests show?
The tests were run on the Weekly part of Memora and on LongMemEval-S. Qwen models in 8B and 30B sizes were used for construction and consolidation. As a comparison, the authors name MemOS, a generative memory system, in each case in the Qwen3-30B configuration.
- Memora Weekly: 46.60 versus 46.68 FAMA (Forgetting-Aware Memory Accuracy).
- LongMemEval-S: 56.8% versus 59.3% accuracy.
- Total LLM token consumption drops by 92.4% and 99.5%, respectively.
What does SLICE-Mem mean for users?
The results suggest that useful cues for a long-term memory can be preserved with significantly fewer model calls. For teams that run agents, this can mean lower costs and shorter waiting times.
At the same time, the comparison on LongMemEval-S shows a small deficit in accuracy. Anyone considering such a method should therefore test it on their own tasks instead of relying only on the benchmark figures.
Frequently asked questions
What is SLICE-Mem?
SLICE-Mem is a lightweight memory framework for LLM agents. It combines an ingestion step guided by the salience of an utterance with an extraction step built around the predicate, the core of a statement.
How does SLICE-Mem work?
A compact classifier decides which user messages are worth storing, and an extractor based on semantic dependencies turns them into concise memory entries. This online pipeline needs no generative LLM calls. Reconciling multiple entries is moved to a later offline consolidation.
What does SLICE-Mem mean for users?
The results suggest that useful cues for a long-term memory can be preserved with significantly fewer model calls, which can mean lower costs and shorter waiting times for teams running agents. On LongMemEval-S, however, accuracy is slightly lower (56.8% versus 59.3%), so the method should be tested on your own tasks.
