Mnema 2.1.0

A password on the memory, from the first start. A new step asks «Protect your memory with a password?». Choose it and the notes and conversations are encrypted on disk with AES-256, the key derived from your password, like KeePass. Type your own password or press «Generate a strong one»; a bar shows how strong it is. Tick «Remember on this computer» and Mnema opens without asking. «No password» stays the default, and the same switch is in Settings → Security.

An encrypted memory stays encrypted when synced. Until now, a memory encrypted by Mnema went to Google Drive, Dropbox, your own folder or a station as plain text. Now only encrypted files leave the computer. A copy encrypted with another password is refused and never overwrites yours.

Meetings, step by step. Settings → Meeting recordings draws the way a meeting goes: Recording → Transcription → Summary → Memory. Each step shows what is set now. Tap a step to see what happens there and its settings: whether to keep the audio, where and how accurately to transcribe, and which model writes the summary. These two choices used to live only in the recording window. Three presets set the whole way at once: Private (everything on the computer), Balanced, and Fast (the cloud Whisper).

One look for every setting. Every choice in the settings is now a small card with its own 3D medal, grey until chosen. That covers learning modes, «What I use to understand texts», gender, themes, how to send, the VPN apps, extra memory folders and the numbers in learning. Switches are tiles. In «What each AI sees» every part of memory has a medal, and «visible / hidden» are real buttons with an eye and a lock. The portrait's actions are tiles too, and «Local memory on your PC» has an open button.

Paid and free, everywhere. The first run, the model menu and Settings → AI & models show AIs in two sections. Paid splits into by subscription and by API key; free means on this computer.

Settings checked under the hood. 48 checks change each setting the way the interface does and look at what actually reaches the model. One fix came out of it: the free model on your computer now also follows the answer length and your gender, not only your name and how to address you.

Smaller fixes.

• The «Input bar» sample is labelled and changes as you flip the switches. The button you just turned on lights up.
• «Pause» during the consortium's last word pauses instead of ending the discussion, and «Continue» lets that AI finish.
• The first-launch splash shows the Mnema M.
• Mnema.exe and mnema-cli.exe carry their name and version.
• The README has a new look, with animations of the real interface.

Mnema 2.0.2

Claude by subscription is found where it actually lives. Mnema used to look for Claude Code only on PATH, and a program started from a shortcut often doesn't see it there. Now it also looks where installers put it — the official installer, npm (including nvm), WinGet, Volta, bun, Scoop, Homebrew — and inside the Claude Desktop app, which brings its own Claude Code.

Found, but not logged in — «Log in». The Claude by subscription tile now tells the two cases apart. If Claude Code is there but nobody is logged in, «Log in» opens it in its own window: log in with your Claude subscription, then press «Check again». If it isn't on the computer at all, «How to install» opens Anthropic's setup page.

Mnema 2.0.1

ChatGPT is easy to find. In Settings → AI & models the tiles are named after the products people look for: ChatGPT (an OpenAI API key) and Claude · API key (Anthropic), next to Claude by subscription and Gemini. The connect window says it plainly: a ChatGPT Plus or Claude Pro subscription gives no API key — the API is paid separately, by use.

The model menu, one rule for every AI. Each connected AI shows its three best models — the newest full-size ones first, then the light ones — and «All models» with a count. Old and special-purpose models (gpt-3.5, gpt-4, Claude 3, dated copies, image, voice and search models) no longer crowd the top: they sit under a line at the end of «All models». The model you picked stays visible even if it isn't among the top three. In «All models» of a connected AI the list comes first, the key and the check fold below.

Every AI can be turned off. Each connected tile has «Turn off», Claude by subscription included. A turned-off AI leaves every list of models — the chat, the consortium, «With and without memory» — while its key and settings stay; «Turn on» brings it back.

An AI that can't answer says why. When Claude's subscription limit is reached, the subscription has ended, there's no money on the account, Claude Code is signed out or a key is refused, the tile shows «not working» with the reason in plain words — for a limit, also when it works again, the AI leaves the lists and the next working one answers. «Check again» sends a tiny real request; a reached limit is tried again by itself in half an hour. The default model is used only while its AI works.

Keys are checked at once. A wrong key is never shown as connected: Gemini's «API key not valid» is now read as a refused key, and an OpenRouter key is checked on its own (its list of models is public, so any key used to pass).

Fixes. Picking «Claude» (the subscription's default model) in the connect window no longer shows «[object Object]». A separate Mnema profile (MNEMA_HOME) offers its own memory folder instead of the main one. The Linux release now carries the .zsync file, so AppImage updaters find new versions.

Mnema 2.0.0

Try it in a minute. Four new things in the «+» menu and on the Memory page. With and without memory: one question to your connected AIs, and each answers twice, as a plain chat that knows nothing about you and through Mnema, side by side. Getting acquainted: seven short questions on day one, and the answers go into memory as ordinary facts without overwriting anything. How well Mnema knows you: a quiz on your own memory, where every answer fixes it (a true status gets today's date, an outdated fact is rewritten, a wrong one goes to the archive). Roast: several AIs take an idea, a resume or a text apart, harsh and to the point, and finish with a score out of 10 and three fixes. Any result becomes a card: a picture with nothing personal beyond what you see on it, to copy, save or post with a ready text.

Consortium, rebuilt. Five formats, each with its own look: council, debate, brainstorm, red team and roast. «Stop» now cuts the AI that is speaking at once and keeps what it said. Every participant gets the whole conversation: the last messages in full, the documents you attached (the passages about the question), the text from your pictures, and the pictures and PDFs themselves. Argue with different AIs under any answer asks who argues; with a single AI connected, its own models argue with each other, for example Sonnet against Opus.

Telegram, read only. Log in with your own api_id and api_hash and pick the chats Mnema may read. In those chats it reads the whole conversation, your messages and the other side's, and learns about you and the people around you. Read only is built into the client itself: sending, editing, deleting, reacting or marking as read is refused before anything leaves your computer. Channels and bots can't be picked. The login is stored encrypted on your computer.

Obsidian and other folders. «Connect Obsidian» finds the vaults on your computer by itself. Any folder can be added as extra memory: a vault is read once in the background and then only the changed notes, and only the passages a question needs go to the AI. Mnema never writes into these folders, and that is enforced in the program: an AI action can't write there even if the folder is open for actions.

Tokens you can see. Under every answer: what it cost (input, cache and answer on hover) and what Mnema saved by reading memory, documents and pages on your computer. The chat list shows both.

One memory switch. The memory is turned on and off in one place, the memory panel on the left. A new memory starts switched off: Mnema is then a plain chat, and the dot by the gear turns yellow. Turning it on always asks first.

Several memories. The settings menu lists the memories on this computer: switch between them, start a new one, rename one or delete it. Deleting asks twice, and the second «Delete» wakes up after ten seconds. The open memory is closed first.

Easier everywhere. The interface starts in English and follows the language you pick everywhere. Mnema's own dialogs replace the browser's pop-ups. The session list keeps the current session on top without empty lines, older ones named by their day, each with a ⋮ menu to rename or delete. One «More precise ↔ Cheaper» slider. Now and then Mnema asks one short question about a project you are talking about (an old status, a missing goal). Nothing is hidden by default: you hide money, health or private topics from an AI in «Who sees what». The interface texts were gone over to be shorter and clearer.

Removed. The browser extension for ChatGPT and Gemini: they work in Mnema's own chat with your key, and whatever the AIs can't do themselves, Mnema does on your computer.

Security. Pages by link and VPN subscriptions open only public internet addresses, never this computer or the local network. The CI token reads only. A VPN key no longer stays on disk after the VPN engine starts; on macOS keys reach the Keychain through standard input, and the secrets file is readable by you only. Regular expressions that could run slow on crafted text were rewritten.

Linux. The AppImage is now Mnema-2.0.0-x86_64.AppImage and carries AppStream metainfo with screenshots of the main window, for app catalogs.

Mnema 1.3.0

Argue with others. Under any answer: its author defends it, one or two other connected AIs look for where it is wrong, and a verdict ends it.

Every model is equal. The model menu and the settings have two groups, cloud and on your computer, each sorted by name. Learning understands your speech with the model you talk to, whatever it is. The answer-length limit applies to every model, Claude Code included.

No network, still an answer. If the internet drops in the middle of a conversation, the model on your computer answers and says so.

Documents on your computer. Scanned PDFs and pictures are recognised locally, so any model can read them. PDFs can be merged, cut, rotated and split. Existing Word and Excel files can be edited in place, keeping their formatting.

Encryption by password, on every system. Notes, conversations and the journal are encrypted on disk (AES-256-GCM, the key protected by scrypt). Mnema can remember the password on this computer or ask for it at every start.

Intel Macs. A separate build for Macs with Intel processors.

System	File
Windows 10 / 11	Mnema-Setup.exe (installer) or Mnema-windows-x64.zip (portable)
macOS 11+	Apple Silicon: Mnema-macos-arm64.dmg · Intel: Mnema-macos-x64.dmg
Linux x86-64	Mnema-linux-x86_64.AppImage or Mnema-linux-x86_64.tar.gz

The same files with the version in the name and their .sha256 are attached as well. The packages are not code-signed yet: see «Install» in the README for the one-time warning on each system.

Mnema 1.2.3

Memory can now be turned off. A new memory starts off and the first tour offers to turn it on. Off means a plain chat: the AIs get nothing from memory and nothing is written; every other function works. One switch in Settings, in the chat's memory panel and in the gear menu.

System	File
Windows 10 / 11	Mnema-Setup.exe (installer) or Mnema-windows-x64.zip (portable)
macOS 11+ (Apple Silicon)	Mnema-macos-arm64.dmg
Linux x86-64	Mnema-linux-x86_64.AppImage or Mnema-linux-x86_64.tar.gz

The same files with the version in the name and their .sha256 are attached as well. The packages are not code-signed yet: see «Install» in the README for the one-time warning on each system.

Mnema 1.2.2

Fixed

• Quitting (the tray's «Выход» or closing the window with «Выйти») could leave Mnema running without a window on macOS. Now it always stops the local model and the VPN engine and exits.

System	File
Windows 10 / 11	Mnema-Setup.exe (installer) or Mnema-windows-x64.zip (portable)
macOS 11+ (Apple Silicon)	Mnema-macos-arm64.dmg
Linux x86-64	Mnema-linux-x86_64.AppImage or Mnema-linux-x86_64.tar.gz

The same files with the version in the name and their .sha256 are attached as well. The packages are not code-signed yet: see «Install» in the README for the one-time warning on each system.

Mnema 1.2.1

Fixed

• «Claude by subscription» showed as connected on computers where Claude Code was installed but nobody was signed in. Now it needs a real sign-in.
• On the first screen the free local model couldn't be picked when Claude was shown. Either AI can be picked now, and a picked local model becomes the default.

System	File
Windows 10 / 11	Mnema-Setup.exe (installer) or Mnema-windows-x64.zip (portable)
macOS 11+ (Apple Silicon)	Mnema-macos-arm64.dmg
Linux x86-64	Mnema-linux-x86_64.AppImage or Mnema-linux-x86_64.tar.gz

The same files with the version in the name and their .sha256 are attached as well. The packages are not code-signed yet: see «Install» in the README for the one-time warning on each system.

Mnema 1.2.0

Windows · macOS · Linux

System	File
Windows 10 / 11	Mnema-Setup.exe (installer) or Mnema-windows-x64.zip (portable)
macOS 11+ (Apple Silicon)	Mnema-macos-arm64.dmg
Linux x86-64	Mnema-linux-x86_64.AppImage or Mnema-linux-x86_64.tar.gz

The same files with the version in the name and their .sha256 are attached as well. The packages are not code-signed yet: see «Install» in the README for the one-time warning on each system.