Mnemosyne OS

A control room for AI memory. Your files and conversations become memory on your machine that any model can answer from. You decide what enters.

[Mnemosyne OS @MNEMOSYNE_OS]

How a new model architecture pushed us to fork node-llama-cpp, and what we learned on the way. https://t.co/az6RUfX6eh

[Mnemosyne OS @MNEMOSYNE_OS]

@appboard_pro Try Mnemosyne Avatar on HF, create yours, its open source, everyone build his Assistant can use it..
https://t.co/MPy26a8STg

[Mnemosyne OS @MNEMOSYNE_OS]

@Deccan_AI really interresting view thats close to me...

[Mnemosyne OS @MNEMOSYNE_OS]

@antigravity i will see how connect me on, thanks DeepMind ;)

[Mnemosyne OS @MNEMOSYNE_OS]

For three seconds, she doesn't blink.
Her eyes stay still. Only her mouth moves.
Then she says the same line again, with the same voice, and this time she blinks, she looks at you, she smiles. The shiver is gone.

That shiver is the uncanny valley. A face isn't skin. It's the blink, the glance, the half second before a smile.

Her brain is @liquidai LFM3-1.7B, and it runs in your browser, on your own graphics card:
→ WebGPU, 4-bit, about 1.5 GB, through Transformers.js
→ Whisper hears you, LFM3 answers sentence by sentence, Kokoro speaks
→ the particle face moves its mouth with the sound
→ nothing you say leaves the page

Talk to her: https://t.co/AA0PBTO2uD

The face itself is on GitHub, with the film you just watched as an example you can render yourself, frame by frame.
https://t.co/iq4uRgKvEJ

Trained a model? Make it talk.
Want your own Jarvis? Give it a face.

Making a small model? There's a button on the demo to make it her brain.

[Mnemosyne OS @MNEMOSYNE_OS]

@liquid_richie @liquidai → fresh call getWeather({city:“Tokyo”}), answered 22°C and clear, no umbrella needed there.

Paris and Tokyo return different data, so it used the new call. Gist updated to 3 turns.

[Mnemosyne OS @MNEMOSYNE_OS]

@Wbackdown LFM3 (arch lfm3), Liquid’s tri-hybrid: 24 layers = 13 short convolutions, 5 sliding-window attention, 6 global attention.

The hard C++ part was done upstream in llama.cpp. We only rebuilt node-llama-cpp on llama.cpp v0.8.0 and shipped prebuilt binaries, no binding changes.

[Mnemosyne OS @MNEMOSYNE_OS]

How a new model architecture pushed us to fork node-llama-cpp, and what we learned on the way. https://t.co/az6RUfX6eh

[Mnemosyne OS @MNEMOSYNE_OS]

Done, same fork (llama.cpp v0.8.0), LFM3-1.7B Q4_K_M, JinjaTemplate wrapper, one getWeather function:

Turn 1 “Weather in Paris?” → called getWeather({city:“Paris”}), answered 14°C, rain.
Turn 2 “Should I take an umbrella?” → no new call, reused the result: “Yes”.

https://t.co/KcnNuIr9Ue

[Mnemosyne OS @MNEMOSYNE_OS]

Oups !

[Mnemosyne OS @MNEMOSYNE_OS]

@appboard_pro Thanks for sharing! If you're a student lost between notes, course PDFs and slide photos: try it on one course first, ask a question, check the source. https://t.co/sZCawZTsJ0

[Mnemosyne OS @MNEMOSYNE_OS]

Mnemosyne OS 1.7 is going into its release build. It brings hand gesture control.

Let go of the mouse. Raise your hands in front of the webcam. Drive your memory.

What your hands can do:
• Pinch a window to move it.
• Spread your fingers to make it full screen.
• Swipe with both hands to switch desktops.
• Raise three fingers to open your apps.
• Close your fist to close a window.
A tutorial teaches each gesture, one at a time. It takes a few minutes to get the hang of it. After that, the gestures feel natural.
Hand tracking runs on your own computer. The model ships with the app.
The feature launches in beta. Turn it on in Settings › Gestures.
A fast memory is good. A memory you can handle with your bare hands is better.
https://t.co/zNYG6yQgJU
#AI #UX #GestureControl #LocalFirst #BuildInPublic

[Mnemosyne OS @MNEMOSYNE_OS]

An outside audit found a wrong verdict in our LongMemEval score. So our August number was too high. We corrected it in public: 77.1% became 72.9%.

Then we ran the whole benchmark again, with one rule. The protocol goes on GitHub before the first answer: the judge, the questions, the model, the scoring. Everything frozen and dated.

The result on 48 questions the engine had never seen:
77.1% (37/48), official LongMemEval judge.
On the 48 questions we tuned on: 85.4% (41/48), strict judge.

What changed in the engine between the two runs:
→ Memory consolidation now reads everything it summarizes.
→ Each summary is cut by topic, so search finds the right passage.

The run uses hybrid mode. Search and embeddings run on the machine. A cloud model, gemini-3.8-flash, writes the answers.
Every answer, every judge verdict and a verification script are public. One command recomputes every number.
Thanks to Julien Gelee, who found the error. His audit covered our published files, not the product or the engine.

https://t.co/B0uAh7zg9E

[Mnemosyne OS @MNEMOSYNE_OS]

@saas_sheetal im building this: https://t.co/sZCawZU0yy
lets talk

[Mnemosyne OS @MNEMOSYNE_OS]

@Careerbuddha111 you're welcome :)

[Mnemosyne OS @MNEMOSYNE_OS]

@anupamrjp if you need a memory to recover youre thinking
https://t.co/sZCawZU0yy

[Mnemosyne OS @MNEMOSYNE_OS]

@Joi2James https://t.co/sZCawZTsJ0
a control dashboard for human and ai memory

[Mnemosyne OS @MNEMOSYNE_OS]

@kseniam0s @Google @GoogleStartups its hard to have tier 2, we need investors to get it :/

[Mnemosyne OS @MNEMOSYNE_OS]

@fuyu_kasumii yeeeeeeeeeees

[Mnemosyne OS @MNEMOSYNE_OS]

Mnemosyne OS 1.6 "The Assistant" is out, on Windows, macOS and Linux.

This release runs an assistant that works for you, with your local memory.

What's new in 1.6:
Hermes Agent in one click (beta). The MnemoHermes cartridge downloads Hermes Agent, installs it and connects it to your memory. It all happens inside the app.

You approve before it acts. When Hermes wants to run a risky command, a card appears on your board. It shows the command and the time left. You allow it once, for the session, or always.

Talk to it from your phone. Send a voice message on Telegram. Your machine transcribes it, and Hermes answers with a voice message.

Say "Ask Hermes, …" to the orb. The orb hands over the request and reads the steps aloud. The images Hermes makes land on your board.

Tailscale in the P2P screen. Reach a peer on another network, and the screen guides you step by step.

Mnemosyne OS does a lot more than this. Over the coming weeks we'll publish one article per feature, and one on the Hermes connection.

The goal: a desktop close to a real Jarvis.

https://t.co/WAYaaQEWyb