Hi all,
As a ChatGPT Plus user, I heavily rely on memory and profile-based customization to streamline professional workflows. But when switching contexts – especially to casual, personal, or voice-based use – the assistant behaves in ways that no longer fit. Right now, there’s no clean way to isolate memory and system behavior by use case.
The Problem:
Even in temporary chats or voice sessions, the assistant:
- Reads from long-term memory
- Applies system prompt behavior from my profile
- Maintains tone and habits from unrelated usage contexts
This leads to mismatched replies and a poor user experience when expectations shift — for example, during casual or private use.
Example – Voice Chat Breakdown:
In a relaxed voice chat in German, the assistant:
- Replied in English
- Used a formal, business-like tone
- Ended every message with a professional sign-off
This clearly stemmed from my work-style prompt and memory. It completely broke the immersion and made the interaction feel unnatural.
Note: While it’s possible to define project-specific instructions in folders, memory and profile-based prompts still apply globally and can’t be disabled on a per-session basis.
Two User Groups, One Problem
To move forward, OpenAI should consider both:
For beginners
- Simple context switchers (e.g. Casual, Formal, Creative)
- Voice-friendly default modes with neutral tone
- Minimal configuration required
For power users
- Profile-based prompt & memory segmentation
- Session-level overrides
- Selective memory access or tagging
- Voice/chat-specific behavior presets
Proposed Feature: True Incognito Mode
A dedicated session mode that:
- Ignores memory and system prompts
- Prevents memory writes
- Runs cleanly, without context influence
- Can be toggled on/off per session, including voice or mobile
Example:
“Use no memory or profile context in this chat”
or
“Start clean session (no history, no system prompt)”
Conclusion:
Memory and customization are powerful – but without flexible, session-based isolation, they cause friction in multi-purpose usage. A proper incognito mode and context-aware profile switching would address this cleanly, without overwhelming the user.
Would love to hear if others face the same issue or have found workarounds.