LycheeMemory V2: Efficient Long-Term Memory for LLM Agents via Semantic Segment-Level Consolidation

TL;DR - LycheeMemory V2 consolidates semantically coherent conversation segments rather than individual turns, reducing the cost of building long-term agent memory while preserving retrieval accuracy.

• Produces context-independent typed memory records from finalized semantic segments.
• Uses lightweight structured indexes and query planning for evidence retrieval.
• Achieves 89.22% on LoCoMo and 92.20% on LongMemEval-S with GPT-4.1-Mini.
• Versus A-Mem, cuts construction tokens by 86.0% and 75.9% on the respective benchmarks without increasing query-time token usage.