Voice-first AI companion Tolan targets natural conversation with GPT-5.1

Tolan, a voice-first artificial intelligence companion built on GPT-5.1, is positioning itself as an alternative to prompt-driven chatbots by prioritising natural, meandering conversation and stable personalities over time.

OpenAI said the app was developed by Portola to support live dialogue that unfolds organically, rather than short question-and-answer exchanges. According to the company, the development effort centred on memory systems and character design, aiming to keep conversations coherent across long sessions and repeated interactions.

Meeting the demands of voice interaction required architectural changes. Tolan reconstructs its context window on every conversational turn, combining a rolling summary of recent dialogue with a persona card, vector-retrieved memories, tone guidance and real-time signals. OpenAI said the introduction of GPT-5.1 alongside its Responses API reduced speech initiation time by more than 0.7 seconds, a meaningful improvement in voice interfaces where latency can disrupt the sense of flow. The system is also designed to handle abrupt topic shifts without losing track of prior context.

For memory and retrieval, Tolan uses the text-embedding-3-large model and stores vectors in Turbopuffer, a high-speed database that supports sub-50 millisecond lookups. The company said it runs nightly compression routines to remove low-value memories and resolve contradictions, reducing noise as conversations accumulate. Personalities are initially seeded by a science fiction writer and then refined by a behavioural researcher, a process intended to balance creativity with psychological consistency.

Portola reported measurable gains following the rollout of GPT-5.1-powered personas. Memory recall misses fell by 30%, while next-day user retention rose by more than 20%. “GPT-5.1 gave us the steerability to finally express the characters we had in mind. It wasn’t just smarter, it was more faithful to the tone and personality we wanted to create,” said Quinten Farmer.

Tolan has attracted more than 200,000 monthly active users since launching in February. Portola said it plans further investment in steerability and memory refinement, including tighter compression, improved retrieval logic and expanded persona tuning. The company framed the work as part of a broader push to show how voice interfaces, backed by large language models, can move beyond commands and prompts towards sustained conversation.

The Recap

• Tolan built a GPT‑5.1 voice companion prioritizing low latency.
• Memory recall misses dropped by 30 percent after upgrade.
• The company plans further steerability and memory refinement.