ASI:One Review: The AI Agent That Remembers Who You Are

The productivity software market already had too many entrants when ASI:One arrived. It’s getting more of them anyway.

$62.5 billion. That’s what the global business productivity software market was worth in 2024, according to market analysis cited by Yahoo Finance. Precedence Research pegged the U.S. portion alone at $17.95 billion that same year. The compound annual growth rate is projected at 14.8% through the rest of the decade. Numbers like that don’t just attract serious builders. They attract everyone, including the ones who’ll post a launch tweet and disappear inside 18 months.

That context isn’t optional. It’s the whole frame for evaluating ASI:One.

The pitch, condensed: a personal AI agent that knows you, remembers you across sessions, talks to other people’s AI agents, and actually executes tasks in the world rather than handing you a numbered list and wishing you luck. Book the reservation. Align the calendars. Handle the follow-through. Automatically. That pitch has been made roughly 40 times in the past three years, and the majority of those products stopped sending newsletters without fanfare.

What’s different here, or what might be different, starts with the infrastructure underneath.

ASI:One is built on top of Fetch.ai, the autonomous agent platform founded by Humayun Sheikh, who also serves as CEO. Fetch.ai hasn’t been building toward this product for a few quarters. It’s been constructing agent-to-agent communication infrastructure for years, and ASI:One functions as the consumer surface of that accumulated work. The Fetch.ai documentation reflects a system that’s considerably more developed than what most consumer AI products are actually running on, which tends to be a wrapper around a frontier model with a clever UI and a memory field the user fills in manually.

The network that ASI:One connects to is called Agentverse, and the team describes it as a marketplace containing millions of agents built to handle research, planning, and real-world task execution. That number deserves scrutiny. “Millions of agents” could mean production-ready tools capable of booking a table or queuing a calendar invite. It could also mean that a significant portion are half-built demos that made it onto the platform and haven’t been touched since. The ASI Alliance’s public roadmap materials don’t resolve that ambiguity cleanly, and it’s worth holding the claim loosely until there’s more independent verification of what the agent inventory actually does in practice.

Still. The scaffolding itself isn’t vaporware. Agentverse exists. People use it. The agent coordination layer is functional. That’s further along than a surprising number of competitors who are promising the same outcomes with nothing comparable underneath them.

The memory layer is where ASI:One’s consumer pitch gets specific. Most AI tools that claim persistent memory make you do the work. You write out your preferences in a settings panel, maybe paste in your schedule constraints, and then hope the system references them consistently. What ASI:One is describing is different in kind: a model of you that builds passively across sessions, accumulating your scheduling preferences, your taste in restaurants, your communication style, your close contacts. Not a profile you fill out. A profile that fills itself out.

That’s the claim, anyway. How well it executes in practice is harder to assess from the outside, and there’s a real difference between a system that learns your patterns well and one that surfaces them at the right moment without becoming presumptuous or intrusive. Those are solvable problems. They’re also not solved problems, and anyone who tells you they’ve cracked persistent memory for a consumer AI agent without caveats is probably glossing over the edge cases.

Planner Mode is the feature that most directly tests whether any of this hangs together.

It’s described as the execution layer for complex, multi-step tasks. The example the team uses is instructive: you say “organize a dinner for eight people Saturday evening,” and instead of receiving a bulleted rundown of things you now have to go do yourself, ASI:One is supposed to decompose that request into subtasks, identify which agents in Agentverse can handle each piece, coordinate between them, and come back with something closer to a completed plan or an actual booking. It’s the difference between a sophisticated autocomplete and something that actually runs errands.

Sheikh has said publicly that “the future of AI belongs to systems,” and while that framing is easy to dismiss as founder positioning, it’s consistent with what Fetch.ai has actually been building toward. Single-model AI assistants that answer questions well are already commoditized. The commercial upside, if it exists, is in agents that can talk to each other, delegate subtasks, and produce real-world outcomes without requiring a human to supervise every handoff.

That thesis isn’t unique to Sheikh. It’s the thesis of most serious AI infrastructure bets right now. The difference is that ASI:One has the full product listing and an actual deployment behind it, rather than a whitepaper and a waitlist.

What remains genuinely uncertain is the gap between feature description and user experience at scale.

Planner Mode working smoothly when an engineer is demoing it and Planner Mode working smoothly when 50,000 users are all asking it to organize dinners, book doctors, and align schedules simultaneously are not the same problem. Agent coordination breaks in interesting ways under load. The “millions of agents” in Agentverse presumably vary enormously in reliability, response time, and capability, and a system that’s supposed to invisibly route tasks to the right agent has to have a good answer for what happens when the right agent fails or returns garbage.

There’s also the interface question. Consumer AI adoption doesn’t hinge purely on capability. It hinges on whether people can figure out what the thing is actually good at and build habits around using it. Planner Mode is a strong feature name. Whether it’s legible enough to a new user who’s skeptical and busy is a separate design challenge.

The market context makes the stakes clear. At $62.5 billion in 2024 and growing at 14.8% annually, productivity software is the kind of category where a well-executed entrant with genuine infrastructure advantages can carve out meaningful territory even without dominating. ASI:One doesn’t need to beat every other AI assistant. It needs to be the one that a specific cohort of users, probably people who’ve grown frustrated by assistants that don’t remember anything and can’t actually do anything, decides to stick with.

Forty prior pitches failed at that. Most of them failed not because the idea was wrong but because the infrastructure wasn’t there to back it up. ASI:One’s argument, implicit in everything about how it’s built, is that Fetch.ai’s years of agent infrastructure work change that equation.

It’s a defensible argument. Whether it’s a correct one will depend on execution that’s still playing out.