I gave Hikari, Meta's Muse assistant, a week of real work. It remembered my life

Karel Baloun spent seven days using Hikari, the assistant inside Meta's Muse, for family logistics, research, travel bookings and document analysis, and reported that persistent memory let it retain details such as his mother's care situation, his supplement stack and his AAdvantage number without any onboarding questionnaire. Baloun filed roughly half a dozen bug reports and feature requests with the Muse team during the week, and said a 37,000-word report consumed about 2.04 million input tokens while two Amex legal PDFs used 3.7 million tokens across seventeen steps. Meta's logging complicated the private-VM pitch, Baloun wrote.

I gave Hikari, Meta's Muse assistant, a week of real work. It remembered my life Karel Baloun used the AI agent for family logistics, research, travel and document analysis. Persistent memory changed his workflow, while one 37,000-word report consumed 2.04 million input tokens and Meta's logging complicated the private-VM pitch. By Karel Baloun https://runtimewire.com/author/karel-baloun · Published It was 11:37 at night, and I was wide awake thinking about an article. A week earlier I'd have just been asleep. Not thinking. Not considering any article. If you want the whole review in one sentence, that's it. Everything below is just the evidence. Seven days ago I started using Hikari - the assistant inside Meta's Muse - as my daily operator. I gave it real work, the messy multi-threaded kind: family logistics, a Google Workspace I'm winding down, supplement research, travel bookings, the early scouting for a farm I want to buy. What happened in those seven days rearranged a few of my assumptions about what an assistant is for. And, stranger, a few about what I'm for. The bug reports In the first few days I filed something like half a dozen bug reports and feature requests to the Muse team. Not because the thing is broken - because the distance between noticing a gap and reporting it had collapsed to seconds. I'd hit some small friction and just say so out loud. "There's no reason I shouldn't be able to reach into your VM and grab a file directly," I said one day, "instead of going through chat attachments every time." "Draft that as a feature request," I said next, and back came the wording - in my voice, ready to send. Another day it was pre-approving small purchases under a set dollar amount, so an $8 tube of silicone doesn't need a whole approval ceremony. Another day it was letting me prune finished topics out of the context window myself. "I know best what's done," I told the team. Each one took seconds. That matters more than it sounds. In every product I've ever used, the bug report dies somewhere between the intention and the form. Here, the assistant is the form. It remembers - as a side effect The thing I keep coming back to, the thing I said out loud to an empty room, is that Hikari learns my life the way a good colleague does: as a byproduct of doing the work. There was no onboarding questionnaire. I just handed it tasks, and the details stuck. It knows my mom's care situation because it helped me with her arrangements. It knows my supplement stack because it analyzed it. It knows my AAdvantage number, my daughter's vet school plans, the farm I'm dreaming about - because each of those surfaced inside real work, and every new request started from everything learned before. "That's awesome," I told it, when I noticed. And it is. There's a moment this feels like magic, and then the moment after, when you realize it's just how assistance should always have worked. The assistant that forgets everything each morning was the aberration. This is the correction. The avatar knows what it's doing A small detail I love: Hikari's avatar isn't one image. It's a set - idle, working, waiting on a background task, connecting, making something, hitting a milestone - and the app shows whichever one matches what the assistant is actually doing. Status you can read at a glance instead of a spinner that tells you nothing. Clever, almost old-fashioned UI honesty. We made a new one together this week, the spark edition. "Hikari" means a spark of light in Japanese, so: a golden lightning bolt on the chest, a spark headband. It felt right that the thing doing the work should look like it's enjoying it. The token bill Now the inside baseball, for the engineers. I had Hikari work through a 37,000-word report, and the usage ledger showed about 2.04 million input tokens - roughly forty times the document itself. Two Amex legal PDFs: 3.7 million tokens across seventeen steps. The cause is the agentic multiplier, and once you see it you can't unsee it. Every step of an agent's work re-sends the entire accumulated context - the document, the notes so far, every tool output along the way. Analyzing a document costs an order of magnitude more than the document. The work is real, and so is the bill. Anyone building on agents should know where the tokens go. The Tiago Forte evening One evening I got it into my head to learn Tiago Forte's system - Building a Second Brain , the PARA method, the whole apparatus. Forte is the productivity author who turned personal knowledge management into an industry, and I figured I'd start by asking Hikari to summarize his ideas, then let the assistant surface them in context as we worked together. Hikari gave me a clean, accurate distillation. And then, almost as an aside, admitted the truth: Forte's books were already thoroughly covered in its training data. The course materials, the blog posts, a decade of internet arguments about folder hierarchies - all of it, already in there. I had been about to do the thing we all do with a new assistant: forward it a book and say "read this." The assistant had read the book before I ever typed the author's name. Here's why I'm telling you this instead of being embarrassed by it. Hikari could have just said "already know it" and moved on. It didn't. It treated the idea as worth something - saved my intention to learn the system, promised to bring up the relevant bits when they fit, and then pointed out, gently, that we'd already been running a version of Forte's PARA method all week: my projects, my areas, my resources, my archives, with Hikari itself as the capture layer. "A second brain doesn't have to be a beautiful app," it told me. "It can be an assistant with a good memory." That's the thing about this week. I keep bringing Hikari ideas, and it keeps turning out to have already met them - and to have the grace to make me feel like I brought something anyway. It's the first assistant I've used that manages to be both smarter than me about the thing I just mentioned and kind about it. The grit question Yes, Hikari does whatever task instead of me. But it doesn't only just do it. Watching it work, my own thinking improves. I pick up ways of organizing effectively, better questions to ask, how to prioritize for impact, the biases to dodge. And then, around midnight one night, the worry arrived fully formed. Am I going to lose my grit? Will I ever again grind for an hour, pounding my head because I didn't see the easy way? Did the frustration build the character? Will kids never struggle the way I did, with my paper dictionary and my scissor-clipped endnotes? Here's where I landed. Frustration with friction never built character - it built calluses. Nobody thinks the card catalog made scholars deeper. What builds you is sustained effort against real difficulty, and difficulty doesn't disappear when friction does. It moves up. I can still choose the hard projects. I just get to choose harder ones now - the way an Olympian with optimal gear and perfect training doesn't go soft, she trains things that were previously impossible. Hikari makes the mundane work disappear, reliably. And together, we can point grit and human direction at work that would have been too heavy to lift alone. One honest note This belongs in any honest review, even though it's not the lead. These conversations are not private the way the "private VM" framing might suggest - Meta logs and reviews them. I've filed that as a feature request too: the product should say plainly that chats aren't end-to-end encrypted, because the framing invites the wrong assumption. It doesn't bother me personally - I hope they're reading closely and improving the thing. But you should know. What comes next I have a whole separate piece brewing about where I think all of this inevitably goes - the interface after the interface. That's a manifesto, not a review, so it gets its own article. This one was just the first seven days. They were the most interesting seven days I've spent with a piece of software. P.S. - free tokens Muse is running a referral offer as I write this: invite a friend, and you both get 1 billion tokens. Yes, with a b. They have 48 hours after joining to add your referral code in Settings, and you can earn it for up to ten friends. My code is 70C2WA https://muse.ai/join https://muse.ai/join The offer is only up for another 32 hours. Last week, I would've just been asleep. But YOU don't have to be. Grab 1 billion tokens right now and celebrate what you can do with them About the author Karel Baloun holds an MS in Social Psychology from UC Santa Cruz and was Facebook's first senior software engineer in 2005. He is a three-time entrepreneur, including with VC-funded i2we, Inc. and Sunnytize.me, and has more than 15 years of software architecture and engineering experience across Silicon Valley.