past.dev's memory API gives AI agents accurate long-term memory, even as facts change over time and many people are involved. It scores 85.03% on BEAM at 10 million tokens, against a previous best of 68.0%.
SAN FRANCISCO, Oct. 6, 2026 /PRNewswire/ -- past.dev today introduced long-term memory for AI agents. Agents that follow a company's email, meetings, messages, tickets and documents need to know which fact is true today, what it replaced, and who may see it. past.dev answers all three, and leads BEAM, the largest public benchmark for agent memory, at every history size.
Most agents still remember by pasting the whole history into the prompt, searching it, or keeping markdown files. None of these knows which fact is still true: in the BEAM paper, models given the full history scored at most 25.9% at 1 million tokens.
How past.dev works
Developers can use the past.dev API to store any timestamped text, then recall it with a question and the asker's identity. past.dev returns the current facts, what they replaced and the dated source behind each, limited to what that person may see, as of now or any past date. It grew out of the agents the team built for companies, where it has run in production for large teams across very large volumes of data.
Results on BEAM
BEAM, published at ICLR 2026, tests AI agents on conversation histories of up to 10 million tokens, beyond almost every model's context window. past.dev ranks first at every size:
History size |
past.dev |
Previous best published* |
100K tokens |
92.08 % |
86.2 % |
500K tokens |
89.63 % |
80.1 % |
1M tokens |
90.65 % |
79.1 % |
10M tokens |
85.03 % |
68.0 % |
"For two years, the industry has tried to solve memory with search. But search finds what looks similar, not what's still true," said Mehdi Djabri, CEO and co-founder of past.dev. "You can rent intelligence. You can't rent a past."
"Deciding that one fact replaced another, across sources, without leaking a private email, is the hard part," said Alex Ronse, CTO and co-founder of past.dev. "We've opened our evaluation harness so anyone can check our work."
Availability
The past.dev Memory API is available today with self-serve sign-up. It works with any model and supports MCP, with self-hosting and dedicated regions on Enterprise. past.dev is SOC 2 Type II audited, ISO 27001 and ISO 27701 certified, and never trains models on customer data.
About past.dev
past.dev (also known as Past AI) is frontier long-term memory for AI agents, co-founded by Mehdi Djabri and Alex Ronse and based in San Francisco. Its memory API tracks what's true now, what it replaced, and who may see it, with a source for every answer. The company is backed by General Catalyst, Varsity and Connect Ventures. Learn more at past.dev.
*Previous best published: the best competing score published as of October 2, 2026, counting only runs on every question at that size with a public, reproducible harness. past.dev's scores use the same publicly available prompts as other top published scores, unchanged, including their guidance from the dataset's question annotations. Full comparison, harness and per-question results: past.dev/benchmarks.
Media contact
Mehdi Djabri • +1 (714) 752-0117 • 424201@email4pr.com
View original content to download multimedia:https://www.prnewswire.com/news-releases/pastdev-introduces-frontier-memory-for-ai-agents-1-on-beam-the-largest-public-memory-benchmark-302899770.html
SOURCE past.dev
