Compact old turns before your context window overflows

A chat that only ever grows. Eventually the model chokes on it.

IntermediateRead 6 minDoing: a few hours8 steps

What you'll have at the end

A chat that keeps answering past the point where its full raw history would have overflowed the window, running on a short summary plus the newest few turns.

Before you start

You need

You have a multi-turn chat or agent loop that already resends its running message history to the model on every call, and a way to count how many tokens that history uses before you send it.

Not covered

Making the assistant remember anything after this chat session ends, or across a brand-new one, is a separate technique with its own persistence layer; this only keeps one already-open conversation inside its token budget while it's still running.

Leans on

8 steps

Where to go next

Chatbot won't drop the old topic

When the growth is really an unrelated new topic still dragging old facts along, not sheer length, reset the topic instead of summarizing it.

Paste only the specific part of a long document that answers the question

When one single huge paste is the growth, not many turns, trim that document before it ever joins the history.

Keep a running scratchpad through one long session

When what's slipping is track of a multi-part task's progress, not the token budget itself, a scratchpad solves that directly.

Checked 15 Aug 2026

Part of the Context Engineering cookbook