Memory extraction can now wait for enough of your own writing

What

Claude Code can pull useful facts out of a conversation and save them as memory for later sessions. The pass that does this now takes an optional config with new settings that control when it runs:

• proseGate (prose_gate): legacy or chat. With chat, the pass is skipped unless you have written enough of your own text, set by minProseWords (min_prose_words), over a reviewWindow (review_window, at least every_n_turns and at least 1). Word counting now handles Chinese, Japanese and Korean text.
• idleFlushMinTurns (idle_flush_min_turns): a minimum number of turns before a save triggered by idle time.
• whenOverage (when_overage): run, first_look or skip, defaulting to run.
• firstLook (first_look): a pre-screen with a mode of off, shadow or enforce, an instruction, a model and max_input_chars.
• The every-N-turns count still reads tengu_bramble_lintel and then defaults to 1.

When the pass is skipped, its log now says "not enough user prose" instead of "no user prose since last extraction". Its usage reports now include the gate in use and a rough word count, and the report for each pass adds fork_model, turns_since_last_pass, prose_gate, prose_words_bucket, user_turns, own_words_turns, own_words_chars, task_notification_turns, unstamped_turns and idle_flush_min_turns.

Why

These settings change when memories are saved and how often the extra work runs. The values come from configuration, not a simple on or off switch, so what you see depends on how it is set.