Memory for AI Agents Gets More Accurate: past.dev Reports the Highest Score Yet on DolphinBench

News provided by

Oct 20, 2026, 08:59 ET

Two weeks after reaching #1 on BEAM, the largest public memory benchmark, past.dev reports the highest score yet on DolphinBench, Mem0's benchmark for AI agents that take real actions.

SAN FRANCISCO, Oct. 20, 2026 /PRNewswire/ -- AI agents are becoming part of our daily lives and work. But as conversations pile up and facts keep changing, most AI agents start acting on outdated or wrong information. past.dev gives agents a memory that navigates the past: what changed, which sources conflict, and where each answer came from. Today it reported the highest score yet on DolphinBench, a new Mem0 test of whether agents act correctly on what users said earlier. Full results: https://past.dev/benchmarks/dolphinbench

past.dev turns a person's or company's chat messages, transcripts, emails and documents into facts an AI agent can rely on. Each fact keeps the date it became true, the source it came from and the fact it replaced. Agents recall what was true for a specific person on a specific date, and each person sees only what they are allowed to see. past.dev ranks #1 on BEAM and reports the highest score yet on DolphinBench. Learn more at https://past.dev

In this example task, a simulated user with nearly four years of history asks an AI agent to correct a pull request about what a December 2023 deploy actually shipped. With past.dev as its memory, the agent gets it right with fewer tool calls and fewer tokens, in less time and at lower cost. Without it, the agent gets it wrong. DolphinBench, Mem0's open benchmark, grades what agents do with a user's history across 600 real tasks. Full results: https://past.dev/benchmarks/dolphinbench

Agent developers and users can use the past.dev API across chat, email, meeting transcripts and documents. Based on public benchmark results, past.dev is the most accurate long-term memory for AI agents available today.

Across 600 simulated tasks in chat, email, calendars, documents and code reviews, DolphinBench grades whether an agent takes the right action based on something the user said earlier.

"An agent can quote a conversation back to you and still send the invoice to the wrong address," said Mehdi Djabri, CEO and co-founder of past.dev. "DolphinBench grades the invoice."

Why memory matters
Models keep getting smarter, but every new LLM session usually starts from zero. As agents like Meta's Muse, OpenAI Dots, Grok Bot, Hark, Instinct and Town become persistent and run for months, their memory will need to get much deeper.

"We're at the very beginning of agent memory," said Djabri. "Soon agents will work alongside us every day, for months and years. Their memory will have to know what changed, why, and who should know. Anyone can rent intelligence. No one can rent a past."

past.dev ran the test itself with Nous Research's Hermes Agent and GPT 5.6 Luna, scoring the highest reported for that setup as of October 20, 2026. Its harness, per-task records and evidence are linked from https://past.dev/benchmarks/dolphinbench. past.dev leads every memory Mem0 tested on that setup, including Mem0, Hindsight, Honcho and Supermemory, on both actions and documents. With its own lightweight agent, past.dev scored even higher, at lower cost and faster than any official DolphinBench result.

"We are asking Mem0 to run us as an official configuration, and until they do our number stays self-reported," said Alex Ronse, CTO and co-founder of past.dev.

past.dev is grateful to Mem0 for an open benchmark anyone can run; competing scores come from Mem0's published runs.

About past.dev
past.dev (also known as Past AI and Past Corp.) builds the most accurate long-term multiplayer memory for AI agents, returning facts with their dates, sources and what they replaced, limited to what each person may see. It works with any model and MCP, and is SOC 2 Type II audited and ISO 27001 and 27701 certified. Based in San Francisco, past.dev is backed by General Catalyst, Varsity and Connect Ventures. Learn more at https://past.dev.

Media contact
Mehdi Djabri
+1 (714) 752-0117
[email protected]

SOURCE past.dev