We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Apr 10·5 min read
Personalizing ChatGPT
TL;DR
ChatGPT users faced inconsistent responses across sessions because the model had no persistent context about individual preferences or requirements. OpenAI built custom instructions and memory features that let users define standing preferences and have ChatGPT retain conversation context across sessions.
✦ Why It Matters
Engineers can now build stateful AI assistants that adapt to user preferences, reducing friction and improving user experience in production systems.
Key Takeaways
- Use custom instructions to define standing preferences once instead of repeating them per conversation.
- Leverage memory features to build context-aware systems that reference past user interactions.
- Design LLM interfaces that persist user preferences to reduce setup overhead.
- Apply personalization at the system level to improve consistency across multiple sessions.
Full Summary
ChatGPT, a large language model (LLM)—a neural network trained on vast text data to generate human-like responses—previously treated each conversation independently, lacking user-specific context. This meant users repeatedly explained preferences, writing style requirements, or domain expertise.
OpenAI introduced two complementary features: custom instructions, which let users specify standing preferences applied to all conversations, and memory, which enables ChatGPT to retain and reference information across separate sessions. Custom instructions function as persistent system prompts, while memory uses retrieval mechanisms to surface relevant past interactions.
These features reduce friction by eliminating repetitive context-setting and enable more personalized outputs. The approach leverages existing LLM capabilities without architectural changes, making deployment straightforward across the platform.