Autonomous work and permissions

Make room with context compaction

A conversation's context holds the recent messages, instructions, tools, relevant memory and other material for the current turn. When it fills up, compaction summarizes the older history so the conversation can continue.

What compaction does

The summary replaces the older messages in what the model reads; your transcript keeps all of them. Compaction never deletes long-term memories.

If a conversation has nothing worth summarizing, compaction does nothing. If it fails, Chetty reports the failure instead of claiming it freed space.

Set automatic compaction

Automatic compaction is on by default and starts at 80% of the model's context window.

Open Account → Models.
Turn Automatically shorten long conversations on or off.
Set Shorten when context is full between 30% and 95%.

The setting also applies to conversations that already exist. Compaction per model sets a separate threshold for one model, because the same percentage is a different amount of text on each.

Note: Conversations in a chat channel always keep automatic compaction, and use the threshold of the account that writes there.

Choose a compaction model

Compaction model in Account → Models picks the model that writes summaries; it can come from a different provider than your chat model, such as a cheaper one.

Left empty, every provider summarizes with the conversation's own model, which can reuse the context the conversation already cached. If your chosen compaction model becomes unavailable, the chat model is used instead.

Compact a conversation now

Send /compact in Web or CLI chat. Add a short note about what the summary must keep:

/compact Keep the deployment decisions and files changed so far.

The conversation shows Conversation compacted when it worked. Otherwise the row explains why, for example Not compacted: nothing to compact yet or Not compacted: already compacted, nothing new since the last summary.

When compaction fails

A failed compaction says so and changes nothing. After several automatic failures in a row, Chetty stops trying automatically in that conversation, and the transcript explains which case it hit:

The summary could not be produced, so automatic compaction is paused for this conversation.
The conversation still measures above its threshold after compacting.

You can run /compact yourself at any time, or start a new conversation, which is the reliable fix for a conversation that cannot shrink.

Admin: Settings → <assistant name> AI → Limits holds Compaction failure limit (1–10, default 3), the number of automatic failures in a row that pauses it, and Compaction summary retries (0–5, default 2), how often a summary request is sent again after a transient error or a refused answer before Chetty falls back to its built-in summarization.

Note: When you return to a conversation whose cached context has provably expired, Chetty may compact it at the start of that turn, but only when compacting costs less than reading the whole history again. Nothing is compacted for a conversation nobody returns to, and never while automatic compaction is off.

Continue with Understand prompt cache alerts, Remove the latest messages with rewind and When a chat grows long or is interrupted.