Esment
← Newsroom
Product update· 3 min read

A free assistant on your own computer, and memory you can fix from the chat

The Esment assistant can now run a small model right in your browser, free and private. Every reply lets you edit or forget the memories behind it, assistants with an API key can change your memory when you ask, and the Assistants page is now Connect, showing which assistants are actually connected.

Markdown

This update is about making your memory easier to use and easier to keep right. You can chat with it without an API key, correct it in the same place you read it, and see at a glance which assistants are really using it.

An assistant that runs on your computer

In the web assistant, choose On this device. An open model, Qwen3 4B, runs inside your browser on your own GPU (2.8 GB, downloaded once), and Settings shows whether your computer runs it well. It costs nothing, nothing is sent to an AI provider, and it answers from your memories like any other model, in a few seconds at most on a modern Mac. It can also save, correct or forget a memory when you ask: every change shows as a card in the chat, with Undo.

  • Esment checks your computer before offering it: it needs a recent browser with GPU access, enough memory and enough free space. On phones, tablets and machines that cannot run it well, the option simply does not appear, so it never slows anything down.
  • It is a basic model for basic jobs: quick answers from what you have saved and short drafts. For long documents or harder questions, OpenAI, Anthropic and OpenRouter are one click away, as before.
  • On the Mac, the built-in local model and Ollama already worked offline. They now get the same memory controls described below.

Fix your memory from the chat

  • Under every reply, the memories it used can now be edited, forgotten or pinned right there, with any model.
  • With OpenAI, Anthropic or OpenRouter, the assistant changes your memory itself when you ask. Say “actually, the meeting is on Wednesdays” and it corrects the memory. This now works in the Mac app too, not only on the web.
  • With a local model, an instruction like “forget the bit about coffee” becomes a short card showing the change. You check it and apply it, and corrections can be undone. Nothing changes until you confirm, so a small model can never get your memory wrong behind your back.

Assistants is now Connect

Assistant (the chat) and Assistants (where you plug in Claude, ChatGPT and others) sat one above the other in the sidebar and were easy to confuse. The second one is now called Connect, and it has been rebuilt around one question: is my assistant connected?

  • Claude, ChatGPT, Perplexity, Grok and Cursor each show whether they have reached your memory and when they last did, based on real connections, not on what you clicked.
  • Connect opens three short steps with a Copy the URL button right next to them. The list updates by itself, so when you finish in the other tab it turns to Connected without a reload.
  • Each assistant links to its own step-by-step guide. Claude, ChatGPT and Perplexity used to point to a general page.
  • The full instructions text, API keys and local setup now sit under Advanced, so the page starts with what most people need.

Fixes

  • Mac: Pin as evergreen in the assistant now pins the memory itself. Before, it saved a pinned copy and left the original as it was.
  • Mac: Forget in the assistant now uses the same soft forget as the dashboard, so the memory stays on its audit trail instead of being deleted outright.

Everything above is live on the web today. The Mac changes arrive with the next Esment for Mac update.

Open the assistant →

— The Esment team

Esment builds persistent, inspectable memory for AI assistants. For press inquiries, write to hello@notas.ai.

← Newsroom