Memo remembers everything, never sends your data to a server, and needs no account or subscription.
curl -fsSL https://download.bugradev.com/get-memo.sh | bashInstall with one command, or grab the AppImage or tarball.
8 GB RAM is enough · no GPU required
This isn't a video — it's Memo's memory pipeline replayed live, right on the page.
A real RAG memory and a tool-calling agent, paired with an interface a first-time user can navigate.
SQLite + sqlite-vec vector search. Every interaction is semantically indexed for O(log n) retrieval. Cross-mode: use a cloud AI for chat while a local model handles embeddings.
27 built-in tools with a sandboxed execution pipeline, up to 40 iterations per turn. File read/write/edit, shell commands, web search and page-fetch, a `change_directory` tool to reach files outside Memo's folder, per-tool timeouts, and a policy-based permission system.
Three gears, cycled with Ctrl+Tab instead of one on/off switch. Plan investigates and writes a saved plan without touching a file, then asks in chat whether to proceed. Auto confirms edits as before. Build runs edits and commands without waiting. Auto-permission chains a finished plan straight into Build, in the same reply.
Hand Memo a `Task.md` checklist and walk away. Worker mode runs each item as its own turn; planner/executor mode plans first and waits for your approval. Up to 3 sub-agents split a large item — one coder, up to 3 parallel read-only reviewers. A busy chat queues instead of dying, a rate limit pauses and resumes from the exact item it was on, and every terminal state notifies you by chat and push.
A small always-on-top character, its own window sharing the same running app, that reflects what Memo is doing right now across chat, WhatsApp, Telegram, and the task loop — thinking, writing, running a named tool, or celebrating a finished turn. Two selectable skins, idle animation, genuinely always-on-top even on Wayland.
How the default setup of each tool stacks up. Products change, so check their sites for the latest.
| Privacy and power, without choosing. | Memo | ChatGPT | Ollama | LM Studio |
|---|---|---|---|---|
| Data stays on your device | Yes | No | Yes | Yes |
| Persistent memory (RAG) | Built in | Limited | No | No |
| Agent tools | Built in | Yes | No | Limited |
| Mobile and chat bots | Yes | Yes | No | No |
| Price | Free | Subscription | Free | Free |
Memo adapts to how you work — whether you are shipping code, writing a thesis, or just want an AI that respects your privacy.
Drop in your codebase. Ask the agent to refactor modules, write tests, or explain a complex function. Orchestra mode splits work across specialist models — one writes React, another writes Go. All offline.
Upload papers, notes, and references. RAG memory connects ideas across weeks of work. Ask "what was that citation about reinforcement learning?" and get the exact paragraph — no keyword search needed.
Your conversations, files, and memories never leave your machine. No account, no cloud dependency. Use external APIs when you need more power — keys are AES-256 encrypted on disk. You hold all the cards.
Windows 10 and 11, Linux (Debian, Ubuntu, Fedora, Arch) and macOS (Apple Silicon and Intel), plus an Android companion app. 8 GB of RAM is enough and a GPU is optional.
Yes. Memo is open source under the GNU AGPL v3. There is no subscription and no account.
Nowhere. Chats, memory and files stay on your device and Memo collects zero telemetry. An external AI provider only sees data if you choose to connect one.
Yes. With the bundled llama.cpp model it runs fully offline. Connect an external provider only when you want more power.
New apps with few downloads can trigger SmartScreen or antivirus warnings. Memo is open source, so you can read the code or build it yourself.
Ollama runs models from a terminal. Memo is a complete desktop app on top of that: persistent vector memory, an agent engine, a mobile companion, and WhatsApp and Telegram bots.
Free, open source and private by construction.
curl -fsSL https://download.bugradev.com/get-memo.sh | bashInstall with one command, or grab the AppImage or tarball.
8 GB RAM is enough · no GPU required