LycheeMemory V2 targets the cost of maintaining long-term memory in long-horizon LLM agents, replacing conventional turn-by-turn memory consolidation with semantic segment-level consolidation that batches multiple interactions before encoding them into structured memory records. The approach reaches state-of-the-art results on two long-term memory benchmarks, scoring 89.22% on LoCoMo and 92.20% on LongMemEval-S. It does so while cutting the number of tokens spent on memory construction by 75.9–86.0% compared to existing systems.
