https://pgmnemo.com/blog/ Measurements, negative results, and engineering notes from building agent memory as a PostgreSQL extension. en Thu, 14 Aug 2026 09:00:00 +0000 https://pgmnemo.com/blog/agent-memory-ab-test https://pgmnemo.com/blog/agent-memory-ab-test Thu, 14 Aug 2026 09:00:00 +0000 Three arms, 435 randomized runs on a production agent fleet, 30 days of real work. Always-on hybrid recall beat the no-memory control on task success (66.9% vs 55.3%, not significant after Bonferroni — every caveat stated). Selective recall, the arm this category sells as intelligence, performed exactly like having no memory at all (p=0.98) despite injecting more context. The de-identified dataset and a standard-library analysis script that reproduces every number are published in the repository.