Problem

The implemented compactor behavior and the public docs disagree.

Current runtime behavior:

• The compactor is context management only.
• It runs a one-turn summarization worker.
• It has no tool server.
• It does not receive memory_save.
• Durable memory extraction is handled by memory persistence branches and explicit memory tools elsewhere.

Stale docs still describe the old model where the compaction agent extracts memories using memory_save.

Files to update

• docs/content/docs/(core)/compaction.mdx
• docs/content/docs/(core)/memory.mdx
• docs/content/docs/(features)/tools.mdx
• src/agent/compactor.rs comments, if still stale
• Design docs only if they are not being changed by an active working-memory PR

Desired invariant

Compaction is context management only.

The compactor watches context size, removes older turns, asks a one-turn LLM worker for a concise summary, and injects that summary into history.

It does not receive memory tools and does not write durable memories.

Durable memory extraction happens through branches, ingestion, cortex, and silent memory persistence branches. Working-memory events provide temporal signals that can trigger persistence earlier.

Suggested verification

rg -n "compactor.*memory_save|memory_save.*compactor|Extract memories|max_turns\\(10\\)|compaction agent has access" docs src prompts
just gate-pr