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

A data note on the dataset Phase-attributed provider-call telemetry from a production agent-memory system (32-day window, 2026-08-07 to 2026-09-08) (10.5281/zenodo.23130191): 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 labelled construction, query or maintenance.

The note reads the table twice. As a cost record: 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% under that filter. As a record of the instrument: 62.8% of query calls match a monitoring or probe signature, so call counts cannot be read as agent demand; the log has a 13-day stretch with no recorded construction or maintenance call and contains no production knowledge-graph extraction. The whole window cost an estimated USD 0.135, so the note reports shares and ratios rather than absolute cost.

Not peer reviewed.

Files

• expensive-writes-noisy-reads-v1.0.pdf: the note (6 pages).
• expensive-writes-noisy-reads-v1.0.md: its source text.
• code-v1.0.zip: analysis.py (downloads the dataset, checks its checksums and recomputes the numbers the note reports), parity.py (binds every number in the note to the metric or dataset-README statement it reports; --selftest checks that seeded errors are caught), figure.py (Figure 1), the PDF build, and the record of the adversarial review the text went through.
• SHA256SUMS.

Reproduce: python3 analysis.py --fetch <dir> --json numbers.json, then python3 parity.py numbers.json expensive-writes-noisy-reads-v1.0.md. Standard library only.