Three arms, 435 randomized runs, 30 days of real work. Always-on recall looked better; selective recall — the part this category sells as intelligence — performed exactly like no memory. Data and analysis script included.
Writing
Measurements, negative results, and engineering notes from building agent memory in Postgres. RSS.
We A/B tested agent memory against no memory at all — on a live fleet. Here is what broke.
2026-08-14