Expensive writes, noisy reads: provider-call telemetry from a production agent-memory system

Characterization studies of agent memory find that building memory costs more than querying it, but they measure benchmark runs, not deployments. We describe a public dataset of 3,544 external model-provider calls recorded over 32 days by the shared store of nox-mem, a self-hosted memory service consulted by six agents. Each call is labelled with a phase: construction, query or maintenance. We read the same table twice. The first reading is about cost. Construction is 3.16% of calls and 55.7% of estimated cost, and one construction call costs about 55 times as much as one query call, or about 25 times once the rows that match a monitoring signature are removed. Construction's share of cost stays between 55.7% and 58.7% whether or not the rows that match a monitoring signature are removed. The second is about the instrument. 62.8% of query calls match a monitoring or probe signature, so the call counts cannot be read as agent demand. The log also has a 13-day stretch with no recorded construction or maintenance call, and it contains no production knowledge-graph extraction. The whole window cost an estimated USD 0.135, which is USD 0.0042 per day, so the note reports shares and ratios rather than absolute cost. Both readings are reported next to the dataset's 16 documented defects, and the code that recomputes the numbers is deposited with the note.