The helper
The helper is an optional download that works quietly in the background so your AIs remember the things you tell them, recall earlier conversations, pick the right model for each question, and make sense of documents you add. It runs privately on this computer - nothing it reads leaves it, even when you chat with online models - and you can remove it any time.
Two small models
| Component | Size | What it does |
|---|---|---|
| Helper model | about 2.2 GB | Remembers what you tell your AIs, grades each question for routing (what kind of task, easy or hard), and makes sense of documents you add. |
| Memory & smart routing | about 334 MB | The memory model, EmbeddingGemma-300M: recalls relevant past conversations, notes and documents, spots by meaning when a question needs the live web or is about your health, and works in more than a hundred languages. |
Both run on their own small servers. After your chat model loads, each starts on the graphics card when the measured room allows, keeping a margin for the chat model, and on your processor when it does not; they give the card back before a model switch, and the decision is taken again once the new model is ready. Settings › Engines has one choice, Automatic or Keep on the processor. On the processor a helper gives its memory back when the computer runs short: an idle helper is stopped under sustained pressure, the activity tray says so, and it starts again on demand once there is room.
The offer
The memory model is offered on its own, once, and the one-time re-index after it lands shows its progress on its card ("Rebuilding memory for your AIs: 3 of 7", "Updating documents"). The helper model is offered on the home page from the second launch only when none of your installed models is graded well enough for memory work: "Add the helper? Your AIs update their memory and read documents better with it." It is never part of the first download, and it waits behind an NVIDIA engine offer, which comes first. Progress survives a page change, and the offer goes away once installed or dismissed.
You can also get either component from Add-ons › Components, and remove them in Settings › Storage - see Components.
What changes without it
Your AIs answer as before. What the helper adds is off, or falls back:
- Memory. Each installed model is graded for memory work by the truth matrix; the model already loaded for chat does the remembering when it meets the grade, and nothing is loaded just for memory work. Without the helper, an AI set to an online model has its facts extracted by that model when no local model is running. Tidying and the picture of you never leave this computer. Recalling earlier conversations and notes by meaning needs the memory model. See Memory.
- Routing. Without the memory model, the live-web and health checks by meaning cannot run: routing keeps ordinary and hard questions on your device and says so, and only a clear live-web cue ("today", "latest", "weather") goes to the search model. Without the helper model, a question's difficulty is unknown: an ordinary question goes to the Everyday online model, and there is no thinking verdict. The offline pick, and Offline Only AIs, work as before. See How routing decides.
- Documents. Making sense of documents you add rides on the chat model instead.