Compaction

When a conversation fills the model's context window, it has to be compacted: the earlier conversation is replaced by a summary. A summary always loses something — exact wording, the reasons behind decisions, instructions from long ago. Fabric layers several things on top of every model's own compaction so that far less is lost, and nothing is lost for good.

The context badge

The percentage next to the model picker shows how full the context window is. Click it for:

You can also type /compact or /compact <what to focus on> in the chat.

What Fabric does

1. Old tool output is cleared first

File reads, command output and screenshots are usually most of the context, and old ones are rarely needed again. Fabric removes older tool output before the model sees it — the newest results and plans/to-dos always stay — so compaction comes much later.

• Claude: Claude Code already clears old tool results on its own and keeps the cleared output on disk.
• Other models (OpenAI, Gemini, Databricks…): Fabric clears old tool output itself just before each model call, once the conversation passes half of the context window. The setting below turns this off.

The agent sees a short note where output was cleared and can run the tool again if it needs it.

2. Recent messages come back word for word

Right after a compaction, Fabric puts the last few exchanges back exactly as they were, so the agent continues mid-thought instead of from a paraphrase.

3. Every summary follows a checklist

Fabric asks for the same sections every time: the goal, your instructions and constraints quoted exactly, decisions and why, the current state, files and identifiers, exact numbers and results, errors and fixes, and the open tasks.

4. Pinned messages are always kept

Hover over one of your messages and click Pin (or the pin icon next to an answer's actions). Pinned messages are restored word for word after every compaction — use it for requirements and rules ("never print secrets", "don't touch the dev instance"). If you pin a great deal (more than about 60,000 characters in all), the later pins are linked instead of repeated, and the agent reads them from the history.

5. The full history stays searchable

Compaction only changes what the model has in view. The full conversation is still saved, and the agent has a search_session_history tool to look up an exact quote, number, path or error from before the compaction instead of guessing.

6. Lasting facts go to project memory

In a project session, when the context gets close to the threshold, the agent is reminded once to add durable knowledge — conventions, decisions and their reasons, preferences — to the project's MEMORY.md, which every future session of the project reads.

7. Undo a compaction

The Conversation compacted divider has an Undo button. It opens a new branch from just before the compaction, with the full context — useful if a summary dropped something important.

8. Compaction waits for a good moment

Fabric compacts after a turn ends, never in the middle of your message. It compacts a little earlier right after a task finishes (a cheap moment) and waits a little longer while background agents are still running (an expensive moment).

Related

• Auto Behaviours — the other things Fabric does automatically
• Conversations — sessions, branching and transcripts