Skip to content

Mnemosyne OS — Infinity Edition v1.3.8 · The Second Channel

Choose a tag to compare

@yaka0007 yaka0007 released this 18 Aug 03:32
· 52 commits to main since this release

What's new

🔎 Your memory now finds the exact word, not just the idea

A question that hinges on a rare literal token — a proper noun, an identifier, a number — could miss the session that contains it: dense embeddings map unique strings close to noise. Retrieval now runs a second, independent channel: your vault is also ranked by exact words (Okapi BM25 over a persistent inverted index, built locally inside the vault database), and the two rankings are merged by rank. Everything stays on your machine — no API, no index service, and hybrid search is ON by default (MNEMO_LEXICAL_FUSION=0 restores the previous behaviour exactly).

Measured on LongMemEval full-haystack with a strict judge: 29 → 37 of 48 questions correct, reproduced across two independent runs — and confirmed on 48 held-out questions never seen during development, with more evidence retrieved and zero regressions. Fusion parameters are literature defaults, deliberately untuned.

🚪 What one app writes, another still cannot read — on every channel

Per-app memory isolation is now enforced inside the retrieval channel itself, not at its call sites — the new lexical channel provably obeys the same ownership rule as the vector channel (5 dedicated tests).

🗂️ A count with an imprecise date says so

Asked "how many X in the last month?", the assistant no longer silently drops an item recorded as just "April 2023". It counts it and tells you the date is imprecise — excluding only what is demonstrably outside the window. The arbitration is yours, not the machine's.

🧠 Local models answer instead of thinking forever

A reasoning model could deliberate past its own reply and deliver nothing. Thinking is now bounded relative to the answer budget: on a 2B reasoning model, 7 empty answers out of 24 runs before, 0 out of 24 after — strongest on modest machines.

🔒 The model's private reasoning stays out of your memory

Streaming a reasoning model's thoughts (so the chat can fold them) no longer lets them reach the internal consumers of the model's return value — capability parsing, vault writes, dream ledgers.

🪟 An auto-opened widget no longer covers the one you opened

Widgets the app opens on its own now stack below the one you are actively using.

Under the hood

The installer now presents the product EULA in English, French and Spanish, and offers to launch Mnemosyne OS when the install finishes. Windows binaries are code-signed (Certum OV certificate, RFC-3161 timestamped), verified on both the Setup and the Portable. A new build gate refuses to sign a package that cannot be imported — the failure class that shipped once can no longer pass. The core also gained exact identity resolution (typed outcomes instead of a lossy null — "we don't hold it" and "you may not read it" are different answers now) with a portable conformance suite.

Downloads

Platform File Notes
Windows Mnemosyne-OS-Infinity-1.3.8-Setup-x64.exe Recommended — auto-updates. Signed.
Windows Mnemosyne-OS-Infinity-1.3.8-Portable-x64.exe No install. Signed.
macOS (Apple Silicon) .dmg Unsigned — right-click → Open on first launch.
Linux .AppImage / .deb x86_64 / amd64.

Existing installs are offered this update automatically.