A memory system earns its keep by reading a conversation once and then answering cheaply forever after. Published figures measure the answering. The reading is an LLM extraction pass over every turn, and it is not free.
Reading one 419-turn conversation cost more than twelve times what every question about it cost put together. One config drives all four systems: same dataset, prompt, answering model, judge and metric.
The crossover below is for one 419-turn conversation asked twenty questions. Yours is a different shape. Put your numbers in and the same measured per-unit costs will tell you where your crossover sits, and whether you are anywhere near it.
Extraction is charged per turn and retrieval per question, so the ratio that matters is questions asked per turn of conversation, not the size of either. A long conversation nobody interrogates is the worst case for a memory system, and a short one asked about constantly is the best.
The per-unit costs are the measured ones: 182.5 extraction tokens a turn, 302 tokens a question against keyword search's 565. What this cannot tell you is whether the answers are good enough, and on this dataset memory scored 0.65 against full context's 0.30, which is the argument that has nothing to do with cost.
Extraction is paid once, before a single question. Retrieval is paid every time. That makes this an amortisation question, and the answer depends entirely on how many questions you are going to ask the same conversation.
Count tokens and the extraction pays for itself after 291 questions. Count dollars, where an output token costs four times an input one, and it is 357. Count only the questions that were answered correctly and it is about 143 solved, which at mem0's accuracy is roughly 221 asked. None of the three is wrong. The accounting choice moves the answer by 60%, which is why the chart above lets you switch it.
Below the crossover, plain keyword search is the cheaper system. That is not an argument against memory; it is an argument that the number belongs on the page next to the accuracy figure.
The same four systems, same questions, same judge. Two of these results argue mem0's case better than its own page does.
Memory beats stuffing the whole conversation in. Full context scored 0.30 against mem0's 0.65, on the same questions with the same model, and it did that while sending 73,892 characters per query. More context made the answers worse, which is the strongest argument for extraction on this page and it is not a cost argument at all.
Retrieval latency is the figure nobody quotes. 76.66ms at p50 against keyword search's 0.97ms, about 79x. It is invisible next to a multi-second generation, and it is not invisible if you are fanning out retrievals.