Rolling Summarization

Replace dropped conversation turns with a condensed summary that is updated each time the window slides. This maintains a coarse history without paying full token cost for every old turn.

How It Works

As the conversation grows beyond the working window, older turns are summarized into a running state block. The prompt sent to the model consists of: system prompt + rolling summary + recent raw turns.

What to Include in the Summary

• current goal
• files touched
• decisions made
• unresolved issues
• next steps

Trade-offs

• Pros: preserves more state than truncation alone; works well for medium-length sessions.
• Cons: summaries can omit nuance; quality depends on the summarization model or prompt.

See Also

• Sliding Window — rolling summarization is typically paired with a working window of recent raw turns.
• Compaction — summarization is the most common compaction strategy in production agents.
• Prompt Summarization — the same technique applied to non-conversation content.

Tools

• LangChain ConversationSummaryMemory — summarizes conversation history to keep context bounded.
• Aider — uses summarization to manage long conversation history in coding sessions.