## v1.2.005 — 2026-06-19
- improvement: **More consistent conversation quality, particularly in non-English languages.** The default sampling configuration shared by local, custom, and cloud models has been retuned from the neutral preset introduced in 1.2.000 to a multilingual-balanced one, which in internal testing across several languages produced more fluent, more coherent replies with fewer stray foreign words. The update applies once to installs that still carry the previous neutral defaults; if you had set your own sampling values, those are left untouched. You can still adjust all sampling parameters in Settings.
- fix: **Long-term memories are recorded in the language of the conversation.** Extracted memories and reflections could come out partly in English during non-English chats. The language instruction for the memory step has been restored, and the conversation's language is detected more accurately, so memories stay in the language you were speaking; proper nouns keep their original form.
- fix: **Internal prompt section labels no longer leak into replies.** On occasion the model could echo an internal marker — such as a memory or character section label — into the chat bubble, the saved history, or the spoken output. These markers were changed to a form the model is far less likely to repeat, and any that still slip through are now removed from the reply, the on-screen bubble, and the text sent to the voice engine.
- fix: **Long conversations keep the character and world settings steady instead of drifting back toward the start.** Once a conversation grew long enough to trim older messages — including after the local model automatically recovers from an interruption — world-book entries injected at a set depth, and a character card's post-history instructions, could be dropped, so the character gradually lost its world setting or persona reinforcement. These are now re-applied after trimming and after recovery, so they stay in context throughout a long session.
- fix: **Recent conversation is retained more reliably in long sessions.** History trimming no longer shortens the kept conversation more than necessary: it preserves a minimum number of recent turns, and its token estimate is more accurate for Chinese, Japanese, and Korean, which were previously over-counted and trimmed earlier than needed.