![](\"https://storage.faqbot.nz/prod/0d5220bc9ad14992d05108dccd460f35/files/pasted_image_article_manage-conversation-compaction_98f3ec11-d0f1-46e3-82e8-62ddcdf0b86f.png\")
Conversation compaction replaces older parts of a long conversation with a summary. This helps keep replies faster and cheaper, while still letting the assistant continue with the important context.
\nIn Theta Assist, you can choose when compaction happens, set the threshold, and review how summaries are created for different model providers.
\nUse these settings to control how long conversations are compacted in Theta Assist. You must be a system admin to access these settings.
\nYou can then choose a compaction mode, set a threshold, and review how summaries are created.
\n\n
The compaction mode setting controls whether users are prompted before compaction happens.
\nWhen a conversation reaches the threshold, interactive threads show a banner asking the user whether to compact the conversation. Scheduled or batch (unattended) jobs compact automatically because nobody is there to respond to the banner.
\nWhen a conversation reaches the threshold, compaction happens automatically without asking.
\nCompaction is turned off. Conversations are not compacted automatically, and the manual Compact conversation action is not available in the thread menu.
\nThe Compaction threshold sets how much of the model's context window can be used before compaction starts or is suggested.
\nThe threshold is measured as a percentage of the current model's context window.
\nFor example, if a model has a 400,000-token context window and the threshold is set to 60%, compaction is triggered at 240,000 input tokens.
\nThere is also a 20,000-token minimum floor, so small-window models are not compacted on every turn.
\nThe Summary creation section explains how summaries are created for different model providers.
\nNewer OpenAI models, including GPT-5.4 and later, support native compaction. These models are marked with a Native compaction badge on the Models page.
\nFor these models, compaction happens on OpenAI's servers using OpenAI's own prompt. The resulting summary can only be read by OpenAI, so the conversation must stay on OpenAI models after compaction. You can't switch to other providers after compaction.
\nOlder OpenAI models do not support native compaction. They use the App summariser instead, like other models.
\nThere is nothing to configure here.
\nClaude models marked as supporting server-side compaction on the Models page are summarised by Anthropic using its default prompt.
\nThese summaries are plain text, so the conversation can still move to other models afterwards.
\nIf a Claude conversation is below Anthropic's minimum size for server-side compaction, the App summariser is used instead.
\nThere is nothing to configure here.
\nModels without provider-side compaction are summarised by the app. An auxiliary model rewrites the older history using the summariser prompt shown on this page.
\nThe summariser model is chosen on the Auxiliary AI Models tab.
\nYou can edit the summariser prompt or restore the default prompt if needed.
\nProviders charge for every token sent to the model. Each new reply re-sends the conversation history, so the cost of a conversation grows as the thread gets longer.
\nOpenAI also charges a higher rate for replies whose input crosses the model's long-context boundary.
\nCompaction keeps the replayed history smaller, which helps keep replies cheaper and below that higher-cost boundary.
\nEach reply reports how many input tokens were used in its final model call. Theta Assist compares that number with the current model's context window from the model catalogue.
\nThis means the threshold is always judged against the model being used at that moment.
\nOnly some models support provider-side compaction.
\nThese capabilities are shown by badges on the Models page.
\nThe outcome is the same either way: older history is replaced by a summary. The main differences are who writes the summary and whether the conversation can move to another provider afterwards.
\nSummaries written by the app or by Anthropic are plain text. This means the conversation can move between providers afterwards, and the new model reads the summary like any other message.
\nThe exception is OpenAI native compaction. OpenAI's summary is encrypted and can only be read by OpenAI, so once one exists, the model picker only offers OpenAI models for that conversation.
\nWhich compaction route is used depends on the model in use when compaction happens.
\nBecause the threshold is measured against the current model's context window, switching to a smaller-window model can make the compaction banner appear sooner.
\nIn Ask before compacting mode, a banner appears when the conversation crosses the threshold.
\nAfter compaction, a small Conversation compacted marker shows where compaction happened. The summary can be viewed from its info icon.
\nOlder messages stay visible on screen. They are simply no longer re-sent to the model.
\nIn Ask before compacting mode, users get a choice on interactive threads only. The conversation is not compacted unless the user confirms it.
\nIn Compact automatically mode, compaction happens silently.
\nA manual Compact conversation action is available in the thread menu unless the mode is Never compact.
\nScheduled and agentic jobs compact automatically in both Ask before compacting and Compact automatically modes. This prevents long-running work from stopping just because nobody is available to respond to a banner.
\nBatch items run as one-shot tasks with no growing conversation history, so compaction does not apply to them.
\nAfter compaction, the model continues from the summary instead of the original older turns. This means fine detail from older messages can drop out of the model's working memory.
\nThe original messages are not deleted and remain visible on screen.
\nMessages with file attachments are never folded away. If a message covered by a summary is deleted, the original messages covered by that summary are brought back into play.
\nUse the summariser prompt to control how app-created summaries are written.
\nUse Auxiliary AI Models (also a tab in the Settings section) to choose the model used by the App summariser:
![](\"https://storage.faqbot.nz/prod/0d5220bc9ad14992d05108dccd460f35/files/pasted_image_article_manage-conversation-compaction_78085301-ee6c-4972-bb70-88617c7d7ac3.png\")
","description":"Learn how conversation compaction works and how to configure it in Theta Assist.","isMarkdown":false,"publishedOnUtc":"2026-09-09T01:43:19.3527623","modifiedOnUtc":"2026-09-08T23:16:51.3924308"}