> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.sparkian.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sparkian.com/_mcp/server.

# Memory and context

AI models don't retain memory between requests on their own. So, with every new message, Sparkian sends along your recent conversation history to keep the reply relevant to what you've already discussed. How much history is included, and in what form, depends on your chat settings and the selected model's context window.

Open **Chat settings** from the **+** tools menu in the message box (or type `/settings`) to adjust this for the current conversation.

## Context mode

Under **Conversation context**, choose how Sparkian handles older messages:

* **Smart** (default) — summarizes older messages, which can use fewer Sparks in long chats
* **Exact** — sends only your recent messages, exactly as written, with no summarizing

## Recent messages

The **Recent messages** slider (0–50, default 20) works a little differently depending on the mode:

* In **Exact** mode, it's a hard cap — only this many previous messages are sent.
* In **Smart** mode, it's the number kept exact; Sparkian can include more than this before summarizing the rest.

> **Note**
>
> Retaining more messages improves continuity but can increase the Spark cost per message, since there's more for the model to read. See [Sparks (credit system)](/sparks).

## Personal memory

Context mode and Recent messages only apply within a single conversation. For Sparkian to remember details about you across separate chats, see [Personal memory](/personal-memory).

## Long conversations

Once an active conversation reaches around **20,000 tokens**, Sparkian shows a notice banner to warn you before costs increase further. Learn more in the [Long context notice](/long-context-notice) guide.