SLICE-Mem preprint reports lower token use for AI agent memory

Researchers report that SLICE-Mem, a framework for building AI agents' long-term memory, reduced total language-model token consumption by 92.4% on Memora Weekly and 99.5% on LongMemEval-S compared with MemOS in tests using Qwen3-30B. A compact classifier selects user messages worth retaining, and an extractor converts them into concise memory entries without generative language-model calls. Reconciliation across entries happens later in an offline consolidation stage. The preprint reports a performance tradeoff: SLICE-Mem scored 46.60 versus MemOS's 46.68 on Memora Weekly's Forgetting-Aware Memory Accuracy measure, and 56.8% versus 59.3% accuracy on LongMemEval-S.