Mengmeng Chen (@mengmengdotchen) header
Mengmeng Chen (@mengmengdotchen) profile picture

Cofounder of @hey_ario - Linking the context that for AI Agents has been missing to be truly useful https://nitter.cf/t.co/kBnYKOGDtw

Palo Alto
Joined January 2011
Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)Photo by Mengmeng Chen (@mengmengdotchen)
It took me about 20 minutes to port over my skills, tasks, and memory to Codex. Within minutes, the AI tool I couldn't imagine working without had been replaced because I just needed to keep moving. Brutal!
1
89
I've gone full Claude Code maximalist for slide making for a while now. HTML output, not PowerPoint. The results are consistently better. Two things bothered me though: HTML files don't go fullscreen easily, and connecting to a projector throws the zoom off mid-presentation. So I spent a few minutes and built a fix. Drop in any HTML file, zoom to fit, go fullscreen. That's it. It's free, nothing leaves your machine. → mengmengchen32.github.io/pre…
133
I've been studying Doubao, ByteDance's answer to ChatGPT. Yes, it has the familiar chat interface. But open it up and you'll find meeting transcription (Otter/Granola), AI podcasts (Notebook LM), PowerPoint creation (Gamma), Homework help (Khanmigo) all in the same app. It's a super app approach to AI. Meanwhile in the US, every function gets its own startup, its own interface, its own login. I think this traces back to user behavior. Chinese users are already trained by WeChat and other super apps to do everything in one place. They don't want to switch contexts. The concentrated tech ecosystem shaped the behavior, and now that behavior is shaping how AI products get built.
114
Why 80% of apps WILL disappear. Ever since I set up my Kaiser skill on @openclaw, I've conversed with the nurse about Ryder's fever, checked in on the baby's four-month checkup, and looked up a health metric from a year ago — all from my phone, using voice input through WhatsApp. No logging in. No navigating. No opening the app at all. I no longer deal with the horrible design that buries the Messaging Center in a place I can never find. Or the tiny text box where I can't see what I've typed. Or the fact that Kaiser buries my health metric inside each "After Visit Summary" — a separate PDF I'd have to download one by one. Not saved anywhere in my profile. No single place to see it. It feels like bliss. "80% of apps are going away" sounded outrageous at first. After this week, I really felt it. Healthcare apps are a particularly extreme example because they're so terribly designed — but the deeper reason is: we were forced to interact with these apps out of pure necessity. Now I don't have to. @steipete
38
This is wild — @openclaw is now my baby's pediatrician liaison: → Nurse messages on Kaiser asking for an update on Ryder's fever → I just say "check what the new Kaiser message is" → Claude Code opens Kaiser via OpenClaw's browser, logs in, switches to baby's profile, reads it back to me → I describe the update in plain English: "milk is back to 700ml, no fever since Friday, sleep regression but we're fine" → Claude Code turns that into a proper reply, I review it, it clicks Send No logging into Kaiser. No navigating that clunky portal. No typing on a tiny phone keyboard. I just talked about my baby like I would to a friend, and a clean medical update went out. thanks @steipete !!
45
The smartest AI model doesn't make the best AI assistant. The one that writes things down and can accurately apply it when it matters does. I gave both ChatGPT's agent mode and my OpenClaw setup the exact same instruction: "Remember that both Ryder (my three-month-old son) and I use Kaiser as our healthcare provider. When I ask you to message his doctor, go there directly." Both confirmed they understood. ChatGPT said "Yes, I will remember that in the future." OpenClaw said nothing special — it just wrote my instruction to a local Memory.md file. Next day. Fresh ChatGPT session. "Can you message Ryder's doctor?" Generic response. No memory of Kaiser. No memory of Ryder's pediatrician. The instruction was gone. OpenClaw loaded that Memory.md file at session start, navigated to Kaiser, found the pediatrician by name, drafted a message, and asked me to confirm before sending. Zero clarification needed. The gap isn't intelligence. ChatGPT's agent runs GPT-5.2 — OpenAI's flagship, one of the most powerful models available. My OpenClaw setup runs Claude Sonnet 4.6, a mid-tier model that's cheaper, faster, and a clear step below by any benchmark. The difference? One agent had a file system it was instructed to write to. The other didn't. It's like hiring two assistants — one writes everything down in a notebook, the other tries to remember. The one with the notebook wins every time. Not because they're smarter, but because the system is better. Agent design works the same way. I didn't even tell OpenClaw to write things down. I said "remember" in plain English, the same way I spoke to ChatGPT. OpenClaw interpreted that as a literal file-writing operation because it had the tools to do so. Simple user experience, meaningful architectural difference underneath. That's what makes context management the real unlock for personalization. A local file system lets the agent build an intelligent hierarchy of what matters — preferences, routines, relationships, shortcuts — and call on it reliably. Not a smarter model. A better memory system.
1
1
46
I wrote about this and the two other hurdles OpenClaw finally cleared in more detail here: mengmengchen.substack.com/p/…
1
27
1/ How does it feel to be claw-pilled? The best way I can describe it is cognitive ecstasy. And it’s extremely addictive. It’s like a flow state — but more explosive, more expansive, more imaginative. I’m building and learning what I’m building at the same time. As I learn more about the infinite possibilities, my list of things to build keeps growing. And because I can’t stop thinking, I can’t stop writing.
1
16
2/ Building feeds learning. Learning feeds imagination. Imagination feeds new ideas. New ideas feed more building. The loop compounds. The closest thing I can compare it to is being around an extremely smart group of colleagues. Except it’s on-demand — no coordination needed — at 2 a.m. — in my pajamas. I am never bored anymore. Waiting in line? Phone comes out. Son’s car nap? Out comes the laptop.
1
13
This morning, my three-month-old had a fever. With a baby in one arm, I voiced my concerns and asked my AI assistant to message the doctor. It navigated to Kaiser, signed in, and sent the message — all on its own. I've been dreaming of this moment for three years. And it actually worked. 🧵
1
31
In the past three years since AI entered our lives, I've always felt — as a founder and a mom — a striking dichotomy between its intellectual capability and its lack of execution. The best researchers at my fingertips. Endless rich conversations. But I'm still the one ordering milk for my toddler between client meetings.
1
15
2025 is the year of AI agents. We want them to feel magical—like they just get it. But if you’re like me, you’re still constantly copy-pasting context just to get them to do the right thing. 🧵 Here's why—and what we're doing about it:
Everyone told us not to open-source this. We’re doing it anyway 🤷‍♀️ Get Gather is a developer tool that connects AI agents to real user data, from Amazon to Zara. It’s the kind of infra that could be its own startup, but we’d rather share it. See it in action on July 9: lu.ma/84lneape Live demos, good people, and a sneak peek at what’s possible when you don’t gatekeep.
7
133
The truth is: AI agents are only as smart as the data you give them. And right now? That data is locked away.
12
Hello AI founders! we are building sth really cool getgather.com but are currently capped by rate limit of OpenAI API. We tried requesting in the portal + contacting sales, all of which we haven’t heard back. Any advice on how to get that solved?
2
1
61
With all the excitement about what we can build with AI and all the world's knowledge that it can be plugged into, we at GATHER feel that a significant piece is missing - your personal context. As smart as ChatGPT is, it does not know anything about you!
1
1
286
At Gather, we are building a set of tools for developers to integrate user data from any b2c system into their apps with permission: playlists from Spotify, Order history from Doordash, Personal bests from Strava and so much more! It's like Plaid for Everything.
1
61
Hello @memdotai 👋 just setting up my Twitter flow. Here's my validation code: zS53BqOpaifrkQ59oSWr
28
Replying to @Airbnb
@Airbnb @bchesky love this travel the world wallpaper in Brian’s guest bedroom! Can possibly to share the link?
1
Replying to @theallinpod
@theallinpod love the pod but every time you guys talk about China, your viewpoints are so simplistic.
The inauguration feels like a group therapy session for the Americans
Truly inspired by female leaders in the VP race. Although it took such a chaos for them to gain the spotlight, this country need their calm and forceful leadership. The sad thing is that they are still sidekicks to the old white male candidate.
I unlocked $49 Rush tickets to #HamiltonSF at @broadwaysf on @TodayTix! Unlock here: todaytix.com/x/sf-bay-area/s…