past.dev (6 Oct) says its long-term memory API for AI agents scores 85.03% on the BEAM benchmark at 10 million tokens, a self-reported result with no independent reproduction, and publishes an open-source evaluation harness. Each stored fact keeps its date, what it replaced and its source, with per-person access.
Sources: PR Newswire: past.dev release (company) · past.dev benchmarks page (company) · Note: The results are self-reported and CTN saw no independent reproduction. “Largest public benchmark” is the company’s description, and the SOC 2 and ISO claims are the company’s.
Read the full story: Agent memory is a data problem. past.dev’s BEAM claim only partly tests it. (Cole Harlan)