How to keep a long session inside the context window

This guide shows how to stop a long session from overflowing the model's context. It assumes you already run mm and know which server serves your model.

Steps

Tell the agent the size of the window. Either set contextWindow in the YAML to the value you started the server with, or leave it at 0 on llama.cpp, where the agent reads n_ctx from /props. The banner shows the result as ctx: <size> (config) or ctx: <size> (/props).
Turn the automatic compression on:
```
context:
  enabled: true
  threshold: 75
  keepLastTurns: 3
```
Before each question, when the history reaches 75 % of the window, the old turns are replaced by one model-written summary and the last three question turns are kept as they are.
Watch the report. With showStats: true (the default) each compression prints one dimmed line:
```
🗜️ compressed 14 messages → 1 summary + 7 kept (18.2k → 4.1k tokens, 6.3s)
```

Variants

The window stays unknown (Docker Model Runner tells nothing, and contextWindow is 0). Only the maxMessages trigger can fire; the agent says so in a start-up warning. Set maxMessages to a message count (one command costs two messages) or set contextWindow.
Compress by hand. Type /compact at the prompt, or send it from the editor in ACP mode. It ignores the threshold but still keeps the last keepLastTurns turns, and says 🗜️ nothing to compact when there is nothing older.
The server was started after the agent. The window is re-probed before the first question while it is still unknown and compression is on; you do not need to restart.
Your own summary prompt. Set context.prompt to replace the built-in seven-section prompt.
A compression that fails (server down, watchdog) leaves the history unchanged and prints [compact: failed, history kept: …].

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37


# How to keep a long session inside the context window

This guide shows how to stop a long session from overflowing the model's context. It assumes you already run `mm` and know which server serves your model.

## Steps

1. Tell the agent the size of the window. Either set `contextWindow` in the YAML to the value you started the server with, or leave it at `0` on llama.cpp, where the agent reads `n_ctx` from `/props`. The banner shows the result as `ctx: <size> (config)` or `ctx: <size> (/props)`.

2. Turn the automatic compression on:

   ```yaml
   context:
     enabled: true
     threshold: 75
     keepLastTurns: 3
   ```

   Before each question, when the history reaches 75 % of the window, the old turns are replaced by one model-written summary and the last three question turns are kept as they are.

3. Watch the report. With `showStats: true` (the default) each compression prints one dimmed line:

   ```
   🗜️ compressed 14 messages → 1 summary + 7 kept (18.2k → 4.1k tokens, 6.3s)
   ```

## Variants

- **The window stays unknown** (Docker Model Runner tells nothing, and `contextWindow` is `0`). Only the `maxMessages` trigger can fire; the agent says so in a start-up warning. Set `maxMessages` to a message count (one command costs two messages) or set `contextWindow`.
- **Compress by hand.** Type `/compact` at the prompt, or send it from the editor in ACP mode. It ignores the threshold but still keeps the last `keepLastTurns` turns, and says `🗜️ nothing to compact` when there is nothing older.
- **The server was started after the agent.** The window is re-probed before the first question while it is still unknown and compression is on; you do not need to restart.
- **Your own summary prompt.** Set `context.prompt` to replace the built-in seven-section prompt.
- **A compression that fails** (server down, watchdog) leaves the history unchanged and prints `[compact: failed, history kept: …]`.

## See also

- Every `context.*` key and its default: [configuration reference](../reference/configuration.md)
- Why the summary keeps recent turns raw and merges earlier summaries: [context compression](../explanation/context-compression.md)