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

# Async Context Compression Filter

> Configure the Async Context Compression Filter, which summarizes and trims long conversations to reduce token usage.

The **Async Context Compression Filter** (`async_context_compression`) is a mAItion **Filter**, not a Workspace Tool. It runs automatically on every message, summarizing and trimming older parts of a long conversation once the token count crosses a threshold, so the model keeps receiving requests it can handle without losing the conversation's context.

## Valves

| Field                          | Required | Default    | Description                                                                                   |
| ------------------------------ | -------- | ---------- | --------------------------------------------------------------------------------------------- |
| `compression_threshold_tokens` | no       | `256000`   | Total conversation token count that triggers compression.                                     |
| `max_context_tokens`           | no       | `512000`   | Hard limit for context. Messages beyond this are trimmed even if compression has already run. |
| `keep_first`                   | no       | `0`        | Number of initial non-system messages to always keep, in addition to all system messages.     |
| `keep_last`                    | no       | `6`        | Number of most recent messages to always keep in full.                                        |
| `summary_model`                | no       | unset      | Model ID used to generate summaries. When unset, uses the current conversation's model.       |
| `max_summary_tokens`           | no       | `64000`    | Maximum length of a generated summary, in tokens.                                             |
| `compression_style`            | no       | `balanced` | Summary compactness: `aggressive`, `balanced`, or `faithful`.                                 |
| `enable_tool_output_trimming`  | no       | `false`    | Whether native tool outputs get trimmed once they exceed `tool_trim_threshold_chars`.         |
| `tool_trim_threshold_chars`    | no       | `10000`    | Character length at which native tool outputs are trimmed.                                    |

All 9 valves in the table above can be edited after install from the filter's settings in mAItion. `compression_threshold_tokens`, `max_context_tokens`, `max_summary_tokens`, `enable_tool_output_trimming`, and `tool_trim_threshold_chars` are additionally set at install time from the env vars below; the other 4 valves start at the filter's own coded default and have no env var.

## Enabling via ENV

The filter can be enabled via ENV. Set in `.env`:

```bash theme={null}
FUNCTION_ASYNC_CONTEXT_COMPRESSION_ENABLED=True
```

`FUNCTION_ASYNC_CONTEXT_COMPRESSION_ENABLED=True` is the only required condition to install. Optionally, set any of the following to override the defaults shown in the Valves table above:

```bash theme={null}
FUNCTION_ASYNC_CONTEXT_COMPRESSION_COMPRESSION_THRESHOLD_TOKENS=256000
FUNCTION_ASYNC_CONTEXT_COMPRESSION_MAX_CONTEXT_TOKENS=512000
FUNCTION_ASYNC_CONTEXT_COMPRESSION_MAX_SUMMARY_TOKENS=64000
FUNCTION_ASYNC_CONTEXT_COMPRESSION_ENABLE_TOOL_OUTPUT_TRIMMING=False
FUNCTION_ASYNC_CONTEXT_COMPRESSION_TOOL_TRIM_THRESHOLD_CHARS=10000
```

Any of these left unset falls back to the default shown in the Valves table above.

The database user mAItion connects with needs permission to create tables for this filter's summaries to persist between requests.

<Note>
  This installs only during first-start initialization — the setup that runs only once, the first time the container boots, gated on a marker file. Setting `FUNCTION_ASYNC_CONTEXT_COMPRESSION_ENABLED=True` on an already-running instance does not install it retroactively.
</Note>