Mnemosyne OS — Infinity Edition v1.3.7 · The Kept Promise

Mnemosyne OS — Infinity Edition v1.3.7 · The Kept Promise

Your memory, your machine. Mnemosyne OS is a sovereign, local-first memory
operating system: your documents, notes and conversations become a living,
searchable memory that never leaves your computer unless you say so.

This release is about the models. The whole loop now runs on your machine
— model, memory and retrieval, with no key and no account. The local
catalogue is current again, with context windows sixty times larger than what
shipped before, and you can load your own. For those who do use the cloud,
providers finally hand you their real model list, and every call keeps a
journal priced the way you price it.

Highlights

🖥️ Fully local, end to end — and genuinely usable

You can now run Mnemosyne OS with no cloud at all: a local model answering
over locally-embedded memory, on your own hardware. No API key, no account, no
call leaving the machine. That was the point of the product from the start; as
of this version it is a working configuration rather than an aspiration.

Said plainly, because you deserve the trade-off and not a slogan: a small local
model recalls less well than a large cloud one. We measure this, and the
gap is real — deep multi-session recall in particular is where a cloud model
still wins. For everyday use — asking your own documents and notes, picking up
a thread, finding what you wrote — the local setup holds up. Start local; reach
for the cloud when a question needs it. Either way the memory stays yours.

🧩 A local catalogue that is current again — and yours to extend

The bundled models had aged between one and two and a half years. Worse, the
small model a modest machine picks first — Phi-3 Mini, 4 GB RAM — had a
4096-token context window, while memory-grounded answers inject far more
than that: the first real question overflowed it. The entry point was the one
guaranteed to disappoint.

Eight new entries, each verified live against the HuggingFace API, all with
a 262k-token context (131k for Gemma):

Model	Size	Note
Qwen 3.5 0.8B / 2B	532 MB · 1.3 GB	runs on very little
Qwen 3 4B Instruct 2507	2.5 GB	no thinking pass — straight answers
Qwen 3.5 4B / 9B	2.7 · 5.7 GB	reasoning
Gemma 3 4B	2.5 GB	a different architecture, on purpose
Qwen 2.5 Coder 3B / 7B	1.9 · 4.7 GB	replaces CodeLlama

Bring your own: point the app at any local .gguf and it appears in the
picker. Nothing was repointed or deleted — models you already downloaded keep
working, marked legacy and moved to the bottom of the list. And local is
now a choice, not a fallback: the picker stops hiding whichever list you are
not on.

🔑 Your provider hands you its own model list

Bring-your-own-key providers now answer the only question that matters: what
can this key actually reach? The list comes from the provider itself, so you
pick from real, current models instead of typing a name and hoping. The key
belongs to the provider rather than to a single model row, a bad key is called
a bad key even when the provider answers 400, and a missing bridge is no
longer reported as an unreachable provider.

💸 A journal of what every call costs — at your prices

Every call now leaves a local journal entry: which model, what it spent.
The app ships no price table — you enter the rates your provider charges
you, and nothing is invented on your behalf. The cost panel shows what the
gate actually debited, a wallet can read its own metered trail, and a
figure that cannot be known shows — rather than a convincing zero.

☁️ Mnemosyne Cloud, in two tiers

For those who would rather not manage keys at all, the sponsored cloud now
comes in two tiers, with the retrieval budget sized to the tier instead of
one setting for everyone.

🔒 Protection you choose, and routing that honours it

Vaults are created normal; you decide what to raise. Marking one
MAXIMUM now genuinely keeps its content off the cloud route — through
ingestion, through the dream layer, through fact extraction, and through the
three text-to-speech paths that could previously read a locally-guarded answer
aloud. The lock is visible before you commit to a conversation rather than
after, and a refused route is a decision you can see and override with “send
anyway”, never a silent wall.

🔄 Updates that tell you what they are

The update dialog now shows the release notes — what changes, before you
decide — instead of a bare Download. On macOS, where builds are not yet
signed and the app cannot update itself, the button opens the release page
instead of doing nothing.

Fixed

• Local models are billed as local: newly added models were loaded on your
  machine and the question still went to the cloud.
• The local model works for a whole session — every second message used to die
  on “No sequences left”.
• Prompts are budgeted against the context window the model actually got,
  not the one advertised, including the first question of a session.
• Reasoning models: their scratchpad is folded on screen and no longer leaks
  elsewhere — and they no longer return an empty answer when a whole reply is
  spent thinking.
• Token estimates were measured on English and were wrong for two other
  languages.
• The assistant no longer invents a shared past with you.
• The chat pill names the model that actually runs, answers name their own
  language, and web search says what is genuinely wrong instead of failing
  quietly.

Downloads

Platform	File
🪟 Windows 11	Mnemosyne-OS-Infinity-1.3.7-Setup-x64.exe (installer) · …-Portable-x64.exe
🍎 macOS (Apple Silicon)	Mnemosyne-OS-Infinity-1.3.7-mac-arm64.dmg
🐧 Linux	Mnemosyne-OS-Infinity-1.3.7-x86_64.AppImage · …-amd64.deb

First-launch notes

• Windows: this build is code-signed (Certum OV certificate, RFC-3161
  timestamped). SmartScreen builds its reputation from download volume, so it
  may still warn on a young signing identity → More info → Run anyway.
• macOS: ad-hoc signed (Apple Silicon only) → right-click the app → Open
  on first launch. A real Apple Developer ID + notarization is coming.
• Linux: chmod +x *.AppImage && ./*.AppImage, or sudo dpkg -i *.deb.

Auto-update is built in: installed apps pick up future releases from this
repository automatically — and, from this version on, say what they contain
first.

---

Everything above runs locally. Cloud features (optional) are strictly
key-gated and never see your vault data.