> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ngram.space/llms.txt
> Use this file to discover all available pages before exploring further.

# You control when it runs

> Pause inference, interrupt a turn, and manage context without losing your Entity.

An Entity can keep working between messages. Its daemon schedules memory work, inner-voice generation, autonomous wakes, and routines. Choose deliberately when that activity should use your model.

## Pause model use

<Tabs>
  <Tab title="Spatial">
    Type `/pause` in the conversation input and select the highlighted command. Use `/resume` to allow model requests again.

    This pauses the shared Entity, including its Telegram conversations and background inference. Other open spatial surfaces refresh their status automatically.
  </Tab>

  <Tab title="Telegram">
    Send `/pause` from a configured operator account. Send `/resume` to continue.

    Operator IDs belong in the private `NGRAM_TELEGRAM_OPERATOR_IDS` environment variable or the Telegram platform configuration. A person who can chat with the Entity is not automatically an operator.
  </Tab>
</Tabs>

Pause cancels active model requests and blocks new chat and embedding requests at the shared inference provider. Scheduled heartbeat, consolidation, and automation entry points skip work while paused. The Entity keeps its identity and memory.

The pause marker, `.inference-paused`, lives beside the Entity's journal. It survives a restart when that directory is persistent. A deployment using an ephemeral disk must first [configure durable storage](/deployment/persistence).

<Note>Pause controls new runtime activity. It cannot reverse a tool action already completed or remove charges for requests a provider has already processed.</Note>

## Stop a response or stop the process

| Action                                   | Effect                                                                                           |
| ---------------------------------------- | ------------------------------------------------------------------------------------------------ |
| **Stop response** or **Esc** in spatial  | Cancels current spatial work and clears response playback. Background operation remains enabled. |
| Spatial `/pause` or Telegram `/pause`    | Blocks model use across the shared Entity until you resume.                                      |
| **Pause creations** in Objects           | Pauses local programs and physics; does not disable background inference or terminate Blender.   |
| **Pause updates** on a Blender object    | Keeps its current preview visible while authoring may continue.                                  |
| **Stop Blender** on a Blender object     | Terminates that project's execution-host process; the last published model remains.              |
| Close a spatial browser tab              | Disconnects that surface. The Entity can continue on other surfaces.                             |
| `/wakequiet on` in Telegram              | Hides wake activity from the chat; it does not pause inference.                                  |
| **Ctrl+C** in a foreground terminal      | Stops that foreground runtime.                                                                   |
| Stop the worker in your hosting platform | Stops the deployed runtime. Local process commands do not stop a remote worker.                  |

For local process cleanup, inspect the targets first:

```bash theme={"theme":"github-light-default"}
uv run ngram stop --dry-run --skip-gateway --leave-ollama-running
```

Then stop those local ngram processes:

```bash theme={"theme":"github-light-default"}
uv run ngram stop --skip-gateway --leave-ollama-running
```

This command operates on matching **local ngram processes**, not one selected Entity. The flags above preserve the separate gateway and Ollama processes.

## Compact without starting over

Live context is the material sent to the model for the next turn. Durable memory is stored outside that context window.

| Control                                          | Behavior                                                                                 |
| ------------------------------------------------ | ---------------------------------------------------------------------------------------- |
| Telegram `/context`                              | Shows the active provider's effective working budget and an approximate usage breakdown. |
| Telegram `/compact status`                       | Shows the current history and rolling-summary footprint.                                 |
| Telegram `/compact`                              | Summarizes older turns, retaining the protected recent turns and a rolling summary.      |
| Spatial `/context` and the context usage control | Inspect context usage and status.                                                        |
| Spatial `/compact`                               | Requests manual compaction of older turns.                                               |
| Telegram `/reset`                                | Clears live conversation history and its rolling summary. Durable memory remains.        |

Compaction uses inference, so resume before requesting it. Short histories may have nothing to compact. If summarization fails, the runtime does not reset the conversation. Automatic compaction reports when it starts, completes, or fails in Telegram and spatial.

The usage display is an estimate, not a provider billing meter. ngram currently assigns the OpenAI Astra route a 256,000-token application working budget. That is runtime policy, not a statement about the provider's maximum context window or prices.

Spatial also provides `/stop`, `/voice`, and `/help`. Commands are highlighted in the composer and handled without a model turn; `/compact` itself uses inference for summarization. [Local creation programs](/spatial/programs) and Blender preview delivery do not initiate per-frame model calls.

## Reduce background usage

Use pause when you need a definite stop. For an Entity that should remain active, tune individual mechanisms in the [configuration reference](/reference/configuration):

* Increase autonomous wake intervals and reduce the hourly cycle cap.
* Disable optional dream generation when you do not need it.
* Review the inner-voice cadence, affect cycle, relational reflection, and consolidation schedule.
* Disable or reschedule routines you no longer use.

`autonomy.enabled: false` disables the autonomous wake engine, but v1 also has a legacy initiative path and independent memory/body jobs. It is not the global inference switch.

For a bounded first experiment, use `uv run ngram ask rook "Hello"`. For a foreground conversation without the daemon, use `uv run ngram talk rook`.