Skip to content
KallamoDocsBack to the site

Long conversations

Every request has a limit. A long chat cannot be sent whole, so Kallamo keeps the recent part active, archives the older part into searchable memory, and recalls it when the story reaches back.

  • Active messages are sent as they are, newest first, as many as fit.
  • Archived messages are split into passages, indexed for search, and summarised by the Summarizer role with a short title and two sentences. They stay readable in the chat and searchable forever.

The original messages are never lost. Summaries are only a map; retrieval searches the messages themselves.

In Configuration › Context & memory:

Setting What it does
Max API payload Everything sent on one request. Retrieved context and older history are trimmed to fit. A smaller limit on the connection wins.
Archive threshold When the active conversation reaches this size, older messages are ready to be summarised. Default 60,000 tokens.
Active conversation size How full the active conversation is, with how many messages are archived, active and dropped. Summarize now archives right away.
Summarize automatically Archive older messages on their own when the threshold is reached.
Context meter in the message box Shows how full the active conversation is, next to the attach button.

If the active history ever exceeds the payload budget, the oldest messages are cut from the request and the chat header tells you, so you can archive them.

The same section lists every summary. You can rename one, rebuild one or the whole archive, finish a summary that was interrupted, or delete a summary and drop its messages.

  • Drop from context: the message stays in the chat, and the AI stops reading it. Useful for a tangent or a mistake you want to keep.
  • Restore to context: the AI reads it again.
  • Revert chat to this message: deletes every message after it. This cannot be undone.

Archiving and the Narrative State work together: what lasts is already recorded as scene, knowledge and relationships, so an archived turn is not needed to remember who knows what. Archived passages still answer questions about how something happened.