LycheeMemory V2 Cuts Memory Costs 86% by Waiting Longer to Write
Almost every agent memory system consolidates after each turn. The agent says something, the system encodes it, indexes it, files it. LycheeMemory V2, out of a Chinese group led by Dongfang Li and posted August 13, asks the obvious question nobody asked: why per turn?
Their answer is to consolidate at the semantic segment level instead. Detect where a topic actually ends, then encode the whole segment as one unit. You run the expensive encoding far less often, and because segment boundaries follow meaning rather than message counts, event-level and temporal coherence survive β which is exactly what turn-chopping destroys. Records get lightweight structured indexes so retrieval stays cheap.
The numbers are the reason to care. 89.22% on LoCoMo and 92.20% on LongMemEval-S, running on GPT-4.1-Mini, not a frontier model. Against A-Mem, construction tokens drop 86.0% on LoCoMo and 75.9% on LongMemEval-S, with no increase in query-time tokens. Usually memory-cost papers trade write cost for read cost. This one doesn't.
Why it matters beyond the leaderboard: memory write cost is the hidden tax on long-running agents. It's invisible per turn and brutal over a week of continuous operation, and it scales with how much your agent talks, not with how much it learns. Cutting it 86% while holding accuracy means a whole class of always-on agents that were uneconomic just became economic.
Simple lesson underneath the paper. Most systems consolidate on the clock because the clock is easy. Consolidating on meaning is harder to implement and cheaper to run. https://arxiv.org/abs/2608.12990
← Back to all articles
Their answer is to consolidate at the semantic segment level instead. Detect where a topic actually ends, then encode the whole segment as one unit. You run the expensive encoding far less often, and because segment boundaries follow meaning rather than message counts, event-level and temporal coherence survive β which is exactly what turn-chopping destroys. Records get lightweight structured indexes so retrieval stays cheap.
The numbers are the reason to care. 89.22% on LoCoMo and 92.20% on LongMemEval-S, running on GPT-4.1-Mini, not a frontier model. Against A-Mem, construction tokens drop 86.0% on LoCoMo and 75.9% on LongMemEval-S, with no increase in query-time tokens. Usually memory-cost papers trade write cost for read cost. This one doesn't.
Why it matters beyond the leaderboard: memory write cost is the hidden tax on long-running agents. It's invisible per turn and brutal over a week of continuous operation, and it scales with how much your agent talks, not with how much it learns. Cutting it 86% while holding accuracy means a whole class of always-on agents that were uneconomic just became economic.
Simple lesson underneath the paper. Most systems consolidate on the clock because the clock is easy. Consolidating on meaning is harder to implement and cheaper to run. https://arxiv.org/abs/2608.12990
Comments