OpenAI

Tolan builds low-latency, personalized voice AI using OpenAI's GPT-5.1


Executive Summary

AI startup Tolan, developed by Portola, has created a voice-first AI companion that engages users in open-ended, natural conversations. By leveraging OpenAI's GPT-5.1 model, Tolan has significantly reduced latency and improved the AI's ability to maintain a consistent personality and context over long dialogues. Their unique architecture rebuilds context every turn, combining memory, persona, and real-time signals to create a more responsive and coherent user experience.

Key Takeaways

* Product: Tolan is a voice-first AI companion app featuring animated characters designed for continuous, evolving conversations rather than simple Q&A.

* Core Technology: The platform is built on OpenAI's GPT-5.1, which provided critical improvements in latency and "steerability," allowing the AI to follow complex personality and tone instructions more faithfully.

* Low-Latency Architecture: Using GPT-5.1 and the Responses API cut speech initiation time by over 0.7 seconds, making interactions feel more natural and conversational.

* Real-Time Context Reconstruction: Instead of caching prompts, Tolan rebuilds its context window every turn by combining a summary of recent messages, a persona card, vector-retrieved memories, and tone guidance. This allows the AI to adapt instantly to abrupt shifts in conversation.

* Advanced Memory System: Tolan uses `text-embedding-3-large` and a high-speed vector database to store and retrieve user facts, preferences, and emotional cues. A nightly compression job prunes low-value memories to maintain quality.

* Measurable Impact: The switch to GPT-5.1 resulted in a 30% drop in memory recall errors and a 20% increase in next-day user retention.

Strategic Importance

This case study highlights the practical advantages of GPT-5.1's improved steerability and latency for complex, real-time applications like voice AI. For Tolan, this technology is a key differentiator, validating their advanced conversational architecture and enabling a more engaging and human-like AI companion.

Original article