We improved conversation memory and stability so the bot can keep track of the conversation, understand context more accurately and respond more consistently in real-world scenarios.
Previously, different conversations could sometimes behave as if they were one large chat. Now communication has a clearer structure: one thread means one character and one channel.
Why: this makes conversations cleaner, reduces confusion and makes it easier to return to a specific topic. It is also the foundation for multiple chats inside the personal dashboard.
Previously, there was a risk of sending too many old messages into the model. That made requests more expensive and could sometimes make responses worse. Now the bot uses only the active part of the conversation: recent messages and only as much context as needed.
Result: replies are more stable and closer to natural conversation — the bot remembers important things without living entirely in the past.
We fixed the order in which modes are applied, for example chat mode or assistant mode. Now the selected mode affects how the bot thinks and responds from the very beginning, not somewhere in the middle of the process.
Why: fewer surprises and more control over what you receive from the bot.
The Web chat is still being prepared, but internally we have already made Telegram and Web work according to the same rules. This is important because there will not be a “second version” of the product that behaves differently.
In live conversations, some words should not change the whole meaning of the dialog. For example, in Rock Paper Scissors, the word “paper” is a move in the game, not a discussion about documents.
This is not only for games. We will gradually add more scenarios to Active Context: trips, mini-plans, support, focus mode and other flows.
