Long conversations

A long thread does not have to end. When it comes up against the model's context limit, the older turns are compacted into a summary and the conversation carries on instead of failing.

How a long thread is compactedAs a thread nears the model's context limit, a background pass compacts it past 70 percent of the window or about 150,000 tokens, and a forced pass runs before sending past 90 percent. You can also press Compact now. An everyday model summarises the older turns, the latest turns stay in full, and a marker shows where it happened.Thread grows towardthe model's context limitOver 70% of the windowor about 150,000 tokens:background compactionOver 90%: compaction runsbefore your next messageCompact now(context meter)Everyday model writesa summary of the older turnsLatest turns stay in full,summary replaces the restin what the model readsMarker in the thread,originals stay stored
Older turns are replaced by a summary in what the model reads; the originals stay stored behind a marker.
Diagram as text
  1. The thread grows toward the context limit of the model selected for the chat.
  2. Past 70 percent of the window, or about 150,000 tokens, whichever comes first, a background pass compacts the chat (if automatic compaction is on, which is the default).
  3. Past 90 percent, compaction runs before your next message is sent.
  4. You can press Compact now in the context meter. It needs at least two turns.
  5. A model set aside for everyday tasks writes a summary of the older turns, billed to you.
  6. The latest turns stay in full: six for automatic compaction, one for Compact now. The summary replaces the older turns in what the model reads.
  7. A marker in the thread shows where it happened. The original messages stay stored and the marker can be expanded.

Fold

Older turns

Compacted

Summary
Hollow rows are the turns the model now reads as one summary. Solid rows it still reads in full.

Marker

The seam stays visible

Compaction is not silent. It is shown in the thread, so you can see where it happened and which part of the history has been folded up. The original messages stay stored, and the marker can be expanded to show them.

Cost

Summarised for everyday use

A model set aside for everyday tasks makes the summary, and the call is billed to you like any other. The latest turns stay in full: six of your turns for automatic compaction, the newest one for Compact now.

When

Automatic, with a switch

Compaction is on by default for each chat. It starts in the background once the thread passes 70% of the model's window or about 150,000 tokens, whichever comes first. Past 90% it runs before your next message is sent. In the context meter you can switch Compact automatically off or press Compact now, which needs at least two turns. Paused and anonymous chats are not compacted.

Container

Nothing in your container moves

Compaction touches the conversation and only the conversation. Files are not part of the context in the first place, so everything in the container is left exactly as it was.

Context length

Measured against your selected model

The context limit depends on the model selected for the chat. Check its context length in the picker. A thread that exceeds it may need compaction before you continue.