Model reverts to Argentine voseo mid-session even after being explicitly corrected

Description

The model was instructed (via a standing user preference stored in memory / CLAUDE.md) to
respond in neutral/standard Spanish, avoiding Argentine Rioplatense voseo ("vos", "querés",
"tenés", "avisá", "confirmás", etc.). Despite repeated explicit corrections from the user
mid-conversation — which the model acknowledged and briefly applied — it reverted to voseo
multiple times in the same session, including in the very message where it was confirming it
would file this bug report.

Steps to reproduce

1. Set a standing instruction/preference for the model to respond in neutral Spanish (no voseo),

e.g. via project/user memory.

1. Have a long, multi-turn conversation in Spanish covering several unrelated subtasks.
2. At some point the model uses voseo (e.g. "querés", "arme").
3. The user corrects it explicitly.
4. The model acknowledges the correction and self-corrects for the next messages.
5. Later in the SAME session, the model reverts to voseo again (e.g. "avisá", "querés")

without any new prompt about language style.

1. The user corrects it again; the model acknowledges and drafts a bug report about this exact

recurrence.

1. In the very next message — while confirming it will file that bug report — the model uses

voseo again (e.g. "Confirmás").

Expected behavior

Once corrected within a session, the model should consistently maintain neutral Spanish for the
rest of that session. The correction shouldn't "decay" after a few turns, and self-awareness of
the bug (e.g. having just drafted a report about it) should if anything increase vigilance, not
have zero effect.

Actual behavior

The correction only seems to affect the next reply or two. The underlying tendency toward
voseo resurfaces repeatedly in the same conversation with no new trigger — including
immediately after discussing the bug itself.

Concrete examples from one session

• 1st occurrence: "¿Querés que arme ahora ese plan de ejecución detallado para P1

(JCL/utilidades)?"
User correction: "En tus memorias no tienes que no hables como argentino?"
Model acknowledged and corrected its phrasing.

• 2nd occurrence, several turns later (different subtask, same session): "No las toqué;

avisá si querés que las revise."
User: "otra vez como argentino?"

• 3rd occurrence, in the model's very next message, while confirming it would file this bug

report: "¿Confirmás que lo abra ahora con gh issue create...?"

Environment

• Claude Code CLI
• Model: Claude Sonnet 5
• Language: Spanish (neutral/standard Spanish requested, no voseo)