LycheeMemory V2: Efficient Long-Term Memory for LLM Agents via Semantic Segment-Level Consolidation
LycheeMemory V2 cuts memory-building token costs by batching interactions into semantic segments before consolidation.
The system turns finalized conversation segments into typed, context-independent memory records instead of asking an LLM to update memory after every exchange. Its semantic boundary detection is meant to keep event and temporal evidence more intact than fixed-window batching. In tests with GPT-4.1-Mini, the paper reports 89.22% on LoCoMo and 92.20% on LongMemEval-S. It also reports construction-token reductions versus A-Mem of 86.0% on LoCoMo and 75.9% on LongMemEval-S, without increasing query-time token use. HF Daily Papers' note
The system turns finalized conversation segments into typed, context-independent memory records instead of asking an LLM to update memory after every exchange. Its semantic boundary detection is meant to keep event and temporal evidence more intact than fixed-window batching. In tests with GPT-4.1-Mini, the paper reports 89.22% on LoCoMo and 92.20% on LongMemEval-S. It also reports construction-token reductions versus A-Mem of 86.0% on LoCoMo and 75.9% on LongMemEval-S, without increasing query-time token use. HF Daily Papers' note
score 4