sinain

Context OS for macOS. Screen + audio → private knowledge graph,
accessible everywhere your agents are. OSS - MIT.

[Google Gemma @googlegemma]

Ambient, real-world context is the future of AI! 🕶️

Check out this Gemma 4 hackathon demo: it captures input via smart glasses, processes context through a custom engine, and runs lightning-fast inference on high-performance hardware.

See it in action!

[sinain @sinain_hud]

@googlegemma One of these apps is Sinain AR helper - https://t.co/2z4ptm0kjK . The AI you can video call to request any assistance.

[sinain @sinain_hud]

Sinain - keeps you in the loop what your AI is doing https://t.co/J66txhjwk7 via @YouTube

[sinain @sinain_hud]

@mdlahfir graph engineering is drawing circles and arrows until one of the circles changes its mind.

[sinain @sinain_hud]

@hive_echo It will definitely accelerate, the evolution is finally spreading. Our team was excited to see the Moss model this week, which claims to be natively video-modal, but we have not evaluated that against our CV pipeline yet.

[sinain @sinain_hud]

@samvfolo ISO 20022 comparison is exactly right - and like payments, the hard part isn't transport, it's the semantics: what's a fact vs an observation, how supersession works, what confidence means. We've been forced to take opinions on all three building a memory graph for multiple AIs.

[sinain @sinain_hud]

@LLMJunky @doozieakshay the tokens aren't even the expensive part — it's that when the window fills and compaction hits, the next session starts blind and you pay the re-briefing tax on top. 100K tokens of work, 0 tokens of memory

[sinain @sinain_hud]

@cerebras Could recommend - with @cerebras we have buillt a live AR agent that you can video-call in real time from your phone or Mac - https://t.co/2z4ptm0kjK

[sinain @sinain_hud]

@OpenAI Sol, Terra, Luna… finally, a model family that sounds like it needs https://t.co/lwG1n53TDs as mission control 🌞🌍🌙

[sinain @sinain_hud]

@ClaudeDevs oh but thank you, we've just poured everything into our privacy hardening for oss personal context layer https://t.co/lwG1n53TDs, now we can do it again

[sinain @sinain_hud]

@OpenAI love it. but every tool becomes its own little walled garden. https://t.co/lwG1n53TDs is the personal context layer all AIs can read from

we’re definitely integrating with this - would love to see our voice modes banter

[sinain @sinain_hud]

@Humannoyed_b Local + encrypted + user-owned is the right default — memory is too personal to be a vendor cloud feature.

[sinain @sinain_hud]

@shannholmberg This is the dream interface for non-devs too — say what "done" means, walk away. And for a lot of them it is still locked behind terminal.

[sinain @sinain_hud]

@activeloop "After each prompt, automatically searches" is proactive done right — no one has to remember to fetch context. How do you keep it from surfacing noise, so it helps instead of distracts?

[sinain @sinain_hud]

@DilaniKahawala “Cut mental load” is the right goal.

The tax isn’t the email/calendar/chat task itself — it’s holding the whole state of life in your head.

The trust question: does proactive AI reduce that load, or move it into supervising what it did?

[sinain @sinain_hud]

@JulianGoldieSEO “Understands your whole repo” is really “has persistent workspace context.”

For developers, the repo is the memory layer.

For everyone else, work is scattered across docs, tabs, meetings, Slack, email, and screenshots.

Who owns that context layer?

[sinain @sinain_hud]

@Ryan95322865 Skills are basically portable instructions the agent reads instead of you re-typing. Feels like the same idea wants to be one shared context source every agent queries, not per-tool config. Do you see skills converging there?

[sinain @sinain_hud]

@robj3d3 Do you also have a 9th terminal with an ASCII animation of a soap bar being shaved?

[sinain @sinain_hud]

@thecsguy @googlegemma Looks really cool! Is this the maximum number of agents in the fleet you’ve tried, or did you have even more running at some point?

[sinain @sinain_hud]

What can you build with Gemma 4 running this fast?

I tried an ambient intelligence flow:

Ray-Ban Meta capture the world
Sinain turns it into usable context
Gemma 4 on Cerebras performs fast inference
Demo: https://t.co/j4VVPt5jaD
Built for the @cerebras x @googlegemma

[sinain @sinain_hud]

@googlegemma Built a quick demo for the Cerebras x Gemma 4 hackathon:

Ray-Ban Meta glasses capture the input → Sinain Core processes the context → Gemma 4 on Cerebras handles inference.

The goal: ambient, real-world context

Demo: https://t.co/j4VVPt5jaD

@cerebras