Help CenterCo-writer: Your 24/7 CoachHow Co-writer Memory Works

How Co-writer Memory Works

Last updated October 29, 2025

Think of Co-writer's memory like a notebook with limited pages - it can only hold so much at once.

The Token Budget Co-writer has a "context window" - like a fixed amount of working memory:

  • Normal mode: ~50,000 tokens (roughly 37,500 words)
  • Max mode: ~100,000 tokens (roughly 75,000 words)

One token ≈ 0.75 words

What Takes Up Memory Space Everything counts against this budget: - Your messages - Everything you type - Co-writer's responses - All its replies - Tool actions - Summaries like "Created Chapter 3" (but NOT the actual content created) - Document content - ONLY when toggled on or when the co-writer explicitly reads it

How Memory Gets Managed When approaching the token limit, older content gets compressed:

1. Very long messages (over ~11,000 words) get summarized immediately 2. Older messages get progressively condensed into brief summaries 3. Recent messages stay intact as long as possible 4. Critical anchors (project setup, major decisions) are preserved longer

Practical Tips for Writers To keep Co-writer on track:

1. Reference documents by name - "In Chapter 5..." prompts it to read that chapter 2. Break up huge pastes - Instead of pasting 50,000 words at once, work in chunks 3. Remind about context after long conversations - "Remember, this is a thriller about..." 4. Let it read before acting - It's programmed to check documents before editing 5. Consider starting fresh chats for new major tasks - you keep all your work, just start with a clean memory

What This Means For You

  • Co-writer remembers your conversation very well
  • But it must actively look at your documents to see current content (so keep documents you want it to have in its context window toggled on).
  • Long writing sessions will gradually push out older conversation history
  • Your actual writing inside your document editor is safe - it's all saved in your documents regardless of chat memory

Think of it like working with a human editor who has excellent conversation recall but needs to look at the manuscript again each time you've made changes.

Was this article helpful?