Data and scripts for: Scaling a local long-term memory server for AI agents from ten thousand to one million records: latency, throughput and CPU cost on one machine

Raw data and scripts for the paper named in the title. The dataset holds the scale series of total-agent-memory 14.3.0, 14.3.1 and 14.4.0 at 10,000, 100,000 and 1,000,000 records (recall and save latency, HTTP throughput with one and four server processes, multi-process writers, database size and memory), five repeated runs at 10,000 records on release 14.6.0 with every latency sample and the host load at each run, per-question paired latency comparisons of search-path changes, vector search tests on real and synthetic embeddings, and CPU probes of the embedding model and the rerankers. analysis.py recomputes every number in the paper from these files; its saved output is included. Two runs taken under heavy background load are kept and marked as excluded. Question ids of LoCoMo (CC BY-NC 4.0) and LongMemEval (MIT) appear in the paired files and stay under those licenses. README.md lists the files.