Tolan: Voice-First AI Companion — Paula Dozsa, Tolan

A voice companion needs to follow a changing conversation without letting old context dominate its next answer. The team behind Tolan describes rebuilding the model's context on every turn, because users interrupt, abandon subjects and return to them while doing other things. Its memory system retrieves relevant information and consolidates duplicates and contradictions overnight. The engineering talk explains this approach.

Reusing unchanged context may preserve a computational cache, but the account identifies a competing requirement: the next response must reflect what the person is discussing now. Memory and immediate conversational context serve different purposes, so the system assembles them again as the exchange develops.