past.dev bets agent builders will pay for memory that knows when a fact went stale

Mehdi Djabri and Alex Ronse built past.dev as the memory behind their own agents for product work, engineering, email and meetings ^[1]. The store keeps each fact with its history, its source and its audience ^[3]. Working out what replaced what has to start from the timestamp attached to each item at ingest ^[4]. The founders list late-arriving information among the problems they ran into ^[14]. A meeting transcript filed under its upload time instead of its meeting time would let an old decision overwrite a newer one.

The access filter has a similar dependency. It acts on the identity the caller passes with each recall request ^[4] and on whatever rights the store holds for that identity. I would ask the vendor two things first: how those rights get into the store, and how quickly a revoked right reaches it. Point-in-time queries add a third: whether a question about an earlier date applies the permissions in force then or now.

The benchmark page describes a single run on September 29, 2026, over complete BEAM splits ^[11]. The 100,000-token split has 400 questions, the 500,000- and 1-million-token splits have 700 apiece, and the 10-million-token split has 200, for 2,000 in total ^[15]. The headline 10-million result comes from the smallest split ^[11]. The scores are not whole-question counts: 92.08% of 400 is 368.32 ^[16]. So the grading is finer than right or wrong. The page also says rival systems were run with different models, judges and configurations ^[13]. Exabase M-1's 68.0% is the next result at 10 million tokens, 17.03 points behind ^[12][17]. For that gap to carry over, one model and one judge would have to score every system. BEAM is a benchmark of long conversations and memory tasks published at ICLR 2026 ^[10]. A team's email and meetings, with an audience attached to every item, is a different workload. I would not treat the score as evidence that the permission filter works.

For a product sold on knowing which version of a fact is current, the launch materials disagree with themselves about the baseline. The release table lists the previous best at 86.2%, 80.1% and 79.1% for 100,000, 500,000 and 1 million tokens ^[18]. Its accompanying graphic shows 76.9% and 71.1% for the first two ^[19]. Measured against the table, past.dev's 92.08% at 100,000 tokens ^[9] leads by 5.88 points. Against the graphic, the lead is 15.18 ^[20]. At 500,000 tokens the lead is 9.53 or 18.53 points ^[21].

Runtime Wire describes the commercial bet this way: teams running agents across internal systems will want one auditable memory layer, so that each application does not have to build and maintain its own history ^[23]. The release says the system ran in production for large teams across very large volumes of data ^[22]. past.dev has not quantified deployments, data volumes, revenue or usage ^[7]. Its About page says early customers get help with evaluation, deployment and production testing ^[8].