글쓰기 스튜디오

Advanced chat: streaming cancel, long-chat compression, book-vs-cross-book memory

약 4분 소요

Subtle behaviours that come up after a while of using the Coach — streaming cancel, long-chat compression, cross-book memory. Here's how they work.

Slima Writing Studio editor as a visual reference: left file tree / centre editor with format toolbar / right Notes panel

1 · Cancel a stream

AI replies stream (typed one chunk at a time) — you can stop mid-stream.

How When
Esc The AI is writing and you see it's going the wrong way
⏸ button next to the message Mouse nearby
Start typing a new message Auto-cancels the current stream

After cancel:

  • What's already produced stays (AI wrote 3 paragraphs, you stopped at 4 → 3 stay)
  • No charge for the remaining tokens (cancelled tokens don't count)

2 · Long-chat auto-compression

When a chat exceeds ~50 messages or ~150K tokens of content, Slima auto-compresses older turns:

  • Early turns → the AI writes a "chat summary" pinned at the top
  • Summary preserves decisions, conclusions, context
  • Original messages aren't deleted, just collapsed — expandable

Trigger notice

The panel shows a banner: "Chat auto-compressed — 12 early messages summarised" + an "Expand original" button.

Avoid being compressed: pin or summarise manually

If you know a chat will be long and want certain turns preserved:

  • Right-click → Pin any important message → won't be compressed (up to 5 pins)
  • Or, at chat end, "Summarize so far" → the AI writes a summary at the top

3 · Book vs cross-book memory

Default: the Coach's memory is per-book:

  • You told the Coach "the protagonist loves the sea" in Book A → Coach in Book B doesn't know
  • Privacy / clean context

Enable cross-book memory

Chat settings (gear) → "Cross-book memory" → pick source books.

With it on:

  • The Coach can read content from other books in the current chat
  • Useful for: screenplay adaptation referring to source novel
  • Or: writing a sequel, staying consistent with the prior book

AI Memory (account-level)

Even more general: AI Memory crosses chats too — your preferences, style, recurring phrasing. Configured in Account → AI Memory.

See: AI Memory: the AI's impression of you

4 · Rate limits / reasoning timeouts

Occasionally you'll see:

  • "Please retry in a moment" → vendor-side rate limit, wait 5–10 seconds
  • "Timed out" → AI thinking exceeded 60 seconds with no response, auto-cancelled

In these cases:

  • Nothing is lost (your prompt remains)
  • Click "Retry" to resend

Related

Was this helpful?