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

# Behavior & memory

> Control how much conversation context the SpeakoFlow assistant sends each turn, whether long chats get summarized instead of dropped, and how long its replies run.

Tune how the assistant carries a conversation, in *Settings → Assistant → Behavior*.

| Setting              | Default     | Options                        |
| -------------------- | ----------- | ------------------------------ |
| Conversation memory  | `12`        | 0 to 200 messages              |
| Summarize long chats | `On`        | On, Off                        |
| Response length      | `No effect` | No effect, Short, Medium, Long |

## Conversation memory

How many recent messages are sent as context each turn. Higher keeps more of the thread in mind at the cost of a bigger prompt. `0` disables it and treats every message as standalone.

## Summarize long chats

On by default. Once the un-summarized part of a conversation grows past your Conversation memory window, the oldest of those messages are folded into a rolling summary instead of being dropped. The six most recent messages stay verbatim.

The summary is injected as a context note ahead of the history, so a long chat keeps its beginning in condensed form rather than forgetting it. Turn it off and a long conversation simply loses its oldest messages once they fall outside the window.

## Response length

`No effect` by default, which adds nothing and leaves your prompt exactly as written. The other three append a length instruction to the system prompt at request time.

Each option tells the model to still match your intent, so a greeting gets a short friendly reply even on `Long` rather than a padded one. Individual [profiles](/docs/personalize/profiles) can override this, and the profile's choice wins for its own turns.

<AccordionGroup>
  <Accordion title="Why some providers get less history than you asked for">
    Cloud providers get exactly the number of messages you ask for. Azure and local engines (the built-in one, or an Ollama or LM Studio server on loopback) also get a character cap on top.

    Azure's gateway rejects oversized request bodies, and a local model runs a small context window, so sending your full history there would fail rather than work slowly. Turns carrying a screenshot trim harder on those same providers, since the image already dominates the budget.
  </Accordion>

  <Accordion title="What triggers a summary, and when it runs">
    The trigger is the message-history window above, not the model's context window.

    The fold runs in the background after a reply finishes, so it never delays an answer, and a subtle indicator shows while it is working. If it fails, the summary is left alone and the next turn tries again.
  </Accordion>
</AccordionGroup>

<Tip>
  You can compact a conversation on demand by typing `/summarize` in the panel. That one replaces the visible transcript with its summary, rather than working quietly in the background.
</Tip>

<Note>
  This is short-term, per-conversation context. For durable facts the assistant remembers across chats, see [Personal memory](/docs/personalize/memory), a separate opt-in feature.
</Note>