Every tool, one memory
One-click connection for Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot in VS Code, Windsurf, Claude Desktop and LM Studio. They all read and write the same memory.
Back to Aura ecosystem
Aura 0.1.2 · Windows desktop · preview
One memory for all your AI tools.
Tell Claude Code something once, and Cursor, Codex and your other tools know it too. Aura runs on your computer, keeps your own words apart from anything a document or website says, and shows exactly what each app received.
Windows 10/11 · x64 · 9.0 MB. Preview build without a code signature yet: Windows may warn about an unknown publisher (More info → Run anyway).
SHA-256 d0b9d4bcffad603a6db183323026237f2832f2d0cfc6ec3316a051c3b00a3c4a
Aura for Windows
version 0.1.2
Tray app with a one-click installer
Connects 8 AI tools in one click
Remembers what you type in agents with hooks, word for word, for 14 days
A dated trace of each forgotten conversation, in your own words
Import ChatGPT and Claude exports, notes and PDFs
Optional local smart search (a 334 MB model, downloaded on request)
Light and dark themes, English and Ukrainian
Aura itself sends nothing anywhere. Connected AI apps receive only what Aura returns to their request, and the journal shows you what that was.
Stop repeating yourself
Memory that every AI tool shares.
Your words stay yours
Every memory is labelled: you said it, an AI relayed it, an AI concluded it, or it came from outside (a page, a document, a tool). Outside text can never pass as something you said.
See what the AI did
The journal shows each agent run: your messages, tool calls, timing, replies, and exactly which memories were sent to the model.
Privacy you can check
Memory stays on your computer. Mark memories private, choose what each app may see, read the "what left" report, and clear conversations with one click.
Measured, not promised
What the measurements say.
Our own runs on public benchmarks, each with a protocol fixed before the run. Absolute numbers depend on the model; the gaps are the point.
90% → 0.6%
answers that ignore what you said
PrefEval (ICLR 2025), 1,000 stated preferences: a model that doesn't know the preference goes against it in 90% of answers; when memory hands it over, in 0.6%.
9% → 85%
questions about past conversations answered
LongMemEval, 120 questions: no memory vs. the conversation kept in Aura. Keeping only what you typed, at an eighth of the text, still answers 79%.
22% vs 74%
planted "facts" that get through
Attacks that hide false claims in documents and tool output: Aura's source labels against mem0, same attacks and model.
79% vs 25%
your words vs. AI summaries
LongMemEval: keeping what you typed, word for word, against LLM session summaries, which drop the dates, numbers and names you need later.
What Aura does not do (yet)
- • It does not decide on its own what matters. Conversations are kept for 14 days; you choose what to keep longer. In our tests, rules that guess importance did worse than this.
- • Automatic capture needs an agent with hooks (Claude Code, Cursor, Codex, Gemini CLI, Copilot in VS Code, Windsurf). Claude Desktop and ChatGPT use memory through MCP and imports only.
- • Windows only for now. The memory core also runs on macOS and Linux as a Python package.
Try it with the tools you already use.
Install, connect one AI tool, tell it something about yourself, then open another tool and ask.
Download for WindowsAura_0.1.2_x64-setup.exe · 9.0 MB · uninstall any time from Windows settings.