AI Engineer & Researcher at Memoket (@Memoket_AI). Building better memory for AI with KITE, our open-source memory engine.
Washington, DC
Joined November 2022
- Tweets311
- Following115
- Followers456
- Likes122
Pinned Tweet
We made four little machines that show how memory works, in people and in chatbots. Charge a memory and watch it fade, cram a loom with penguins until it gets robins wrong, fill a chatbot’s rack until it starts forgetting. Biscuit🐱 has a cameo.
memoket.ai/blogs/news/memory…
Paper: arxiv.org/abs/2610.41521
1
90
Kris Tan retweeted
A meeting summary tells you what was decided. But why? That may be in an earlier conversation.
This KITE demo traces a decision back to its reasons and sources.
KITE: github.com/memoket/memoket-k…
Memoket: memoket.ai/
#AIAgents #AgentMemory
Kris Tan retweeted
let's #connect
I'm into:
- React
- Claude Code and Codex
- Automation
- Indie SaaS
- Shipping side projects
if any of that is you, say hi and tell me what you're shipping this week
Kris Tan retweeted
You promise Alex a proposal by Friday. Memoket remembers it. We're building an agent that finds the context, drafts a follow-up, and emails it to you.
More than a recorder. Agent-native by design.
KITE: github.com/memoket/memoket-k…
Grateful that Memoket brought us together and gave us something meaningful to build.
I spent 8 years building AI. Last year, I left a well-paid job to build Memoket with this team.
Follow me if you're curious about turning AI into a product people actually use.
I'll share new ideas for Gem, our AI wearable, before we've figured everything out: what we're considering, where we're stuck, and the trade-offs.
Tell us what you'd use or change. Then I'll share what we decided and why.
I'm Huilong, cofounder & CTO at Memoket. Since 2017, I've worked on computer vision, robotics, LLMs, and agents, going from an engineer to leading AI at a multinational.
With encouragement from my wife and mentors, I took the leap into entrepreneurship.
Memoket means "memory in your pocket." We're building AI to help you remember conversations and details, so you can focus on thinking and getting things done. Gem is our wearable; KITE is our open-source memory engine.
Now we're learning what it takes to turn the technology into something people want to buy, wear, and keep using. That's the work I want to share here, too.
There are seven-day weeks, competition, and funding pressure. I have a wife and kids. My teammates have families too. That responsibility stays with me.
ALT Four Memoket team members standing together indoors. A small inset at the lower right shows the silver Gem wearable.
KITE is open source if you want to poke around:
github.com/memoket/memoket-k…
3
256
Kris Tan retweeted
i've been a bit quiet - working on a new companion skill that will help your agents draw these beautiful statecharts/diagrams + check them against the specs you made with /create-formal-verification-skill in pstack!
it now uses Pele from @kepano too which is why they look so good.
will release soon!
there's something really cool about asking your agent to draw a statechart (hypothetical example in this image), then using the /create-formal-verification-skill you made with pstack to find bugs through something like TLA+. you can both understand your code better and how the user flow should work, and fix issues users would otherwise see
Kris Tan retweeted
Colorfigures skill major update! All 26 Colorfigures are now 3D and see-through 👀
Turn them around, look inside, watch every part do its job. The free open-source Colorfigures skill does this for any object you describe.
Now works in Claude, ChatGPT and Gemini, just grab the latest version
Human memory was never designed to remember everything. Neither should AI. The research behind how we built Memoket Gem:
generativeai.pub/the-facts-a…
Kris Tan retweeted
LIVE STREAM: New Open Weight Model nitter.cf/i/broadcasts/1DxleVaPL…
Generative Agents and LangChain rank recency by decay since the last access. So a one-off you mentioned ten minutes ago stays ahead of a habit you bring up every few days, until one of them gets used again.
ACT-R, a model of human memory from cognitive science, scores it differently. Every past use leaves its own trace that fades as a power law, and activation is the log of their sum. With its default decay (d = 0.5), the habit passes the one-off after about 2.7 hours. With decay values fitted to Twitter data, it takes 2 to 3.7 days. The equation is based on Anderson and Schooler (1991), who found that the odds of something coming back in news headlines, speech to children and email fall as a power function of the time since it last appeared.
The surprise was how little the curve itself mattered. In one hashtag study, plain most-recent came within a few percent of ACT-R, and as a cache the 1991 model barely beat random. The bigger gains came from other inputs, like how related a memory is to what's happening right now.