Fast decisions

Does a conversation contain something worth remembering? Does a new memory repeat something already saved? What does “last month” mean in a search? Fast decisions lets Mem use a dedicated decision model for these small judgments. Your chat model continues to write answers and extract memories.

This setting is optional. You can keep using Mem without it. To try it, open Settings → Providers → Fast decisions.

Turn it on

If the row already shows Nowledge AI, your account has configured it for you. You do not need another API key. To use your own provider instead, click Use my own provider, then follow the steps below.

Fast decisions requires the right-hand option in the model switch above the assignments: Remote, or Nowledge AI when your chat uses it. Changing this mode also affects the models used for chat and background work. Local mode pauses Fast decisions, even when a provider is saved. If you want to keep using only the built-in local model, leave this optional feature off.

You can connect directly to TypeSafe AI, or use an OpenRouter, Vercel AI Gateway, or Cloudflare AI Gateway account. If you already use OpenRouter or Vercel for chat in Mem, select that saved connection in the Fast decisions dropdown and click Test. Mem uses Jev for decisions while keeping your chat model selection.

To add a connection:

1. Get an API key from your chosen provider. For a direct connection, use the TypeSafe AI console. A gateway connection uses that gateway’s key; it does not need a separate TypeSafe key.
2. In Mem, click Set up beside Fast decisions, then choose the provider. If Set up opens TypeSafe AI directly without a menu, update the app to add gateway choices.
3. Enter the key and keep the suggested Jev model. For Cloudflare, also fill in your Account ID; see the details below.
4. Click Verify. A successful check also saves the connection. If you make further changes, click Save.
5. Confirm that the model switch above is on the right-hand option, then select your provider in the Fast decisions dropdown. The row now shows your provider and Jev model, and the selection takes effect immediately.

Saving the provider stores its connection details. Step 5 assigns it to Fast decisions. Jev answers decision questions and cannot be used as a chat or vision model. OpenRouter and Vercel can still supply other models for those tasks.

Cloudflare account details

Use a Cloudflare API token with Account → Workers AI → Read permission for the account you want to use. Copy its 32-character Account ID from the Cloudflare dashboard into Mem. Gateway ID is optional: enter an existing gateway’s ID to send requests through it, or leave it blank to use your account’s default gateway. See Cloudflare’s Jev guide for account-side setup.

Which Jev model to use

Mem selects the model for the provider you choose. The names differ between providers:

OpenRouter’s typesafe/jev-router is a different model and is not used for Fast decisions. You do not need to replace your saved chat model with Jev.

Check the connection

For your own provider, click Test beside Fast decisions. A result such as “Answered in … ms” confirms that the selected model responded. The test sends a short example question, which your provider may charge for. A Nowledge AI row has no Test button and needs no manual connection test.

That time measures one test call. It is not the time taken by memory retrieval or a complete Deep Search.

Once enabled, Mem uses the model in supported steps, including screening conversations for extraction, checking for duplicate memories, and interpreting time references in Deep Search. Ordinary fast search does not make extra decision-model calls by default. Time ranges inferred during Deep Search help rank results; explicit date filters remain under your control.

Where your data goes

The conversation excerpts, memory content, or search queries needed for a judgment go to your selected provider. A direct TypeSafe connection sends them to TypeSafe AI. A gateway connection sends them through OpenRouter, Vercel, or Cloudflare to TypeSafe AI. Billing follows your gateway account’s configuration, and its data-handling settings also apply.

These calls still leave your device if Default or Agents uses LM Studio or Ollama. To pause them, turn Fast decisions off or switch to the built-in Local mode.

To keep all model calls on your device, leave Fast decisions off and check the other settings in the fully local models guide.

For Nowledge AI, use the configuration shown by your account. These gateway choices are available in Vault App. In hosted Cloud workspaces, Nowledge manages the decision model using TypeSafe’s official service. Cloud does not offer gateway selection or this Providers screen.

Turn it off or troubleshoot

To disable your own configuration, select Off in the Fast decisions dropdown (Not configured in older versions). The change saves automatically. Your existing memories, chat model, and saved provider API key remain in place. If the row shows Nowledge AI, click Use my own provider first, then leave Fast decisions off.

Switching to Local mode also pauses Fast decisions while preserving its configuration. The selected decision model resumes when you switch back to Remote or Nowledge AI.

If your provider is missing from the dropdown, check that you have verified and saved its connection. Entering an API key alone does not complete setup. For Cloudflare, also check that the Account ID belongs to an account your token can access.

If the test fails, check your API key, available provider credit, and network connection, then test again. You can return here to check the connection after setup, too.

Mem also checks the Jev version returned by a gateway. If the version is missing or not yet supported, verification fails and Mem keeps its usual workflow without those judgments. You can use a direct TypeSafe AI connection while waiting for support for that gateway’s response.

See LLM Providers for your chat, agent, and vision model settings.