🧠 The local model gives your memory back#
Until now, the first thing that touched local AI loaded the model and nothing ever unloaded it. One Smart Action at 9am and several gigabytes stayed held until you quit the app. The model now releases itself after a few minutes without use, and there is a small Local AI marker at the top of the window telling you exactly how much memory it is holding, with a Free memory now button when you want it back immediately. If you would rather never pay the wait, a new Keep the local AI model ready setting loads the model at startup and holds it for the whole session instead.
⚡ The first local answer lands in about a second#
The first message of a chat session used to take around 26 seconds before anything appeared, because the model had to read your system prompt and every tool definition before it could start. Azynote now does that work ahead of time, when you open chat or the Smart Actions panel. Same model, same machine, about 1.5 seconds to the first response.
🤝 One Azynote no longer stops another's model#
If you had Claude Desktop or Cursor connected to Azynote, a single search from them could kill the model you were mid-conversation with, and your next message paid the full cold start again. Azynote now tracks which process owns which model, so cleaning up after a crash only touches models nobody is using.
📝 OnePagers always read the cleaned-up transcript#
A OnePager could be built straight from raw transcription output, with quotes attributed to "Speaker 2" and uncorrected text, if you forgot to hit Improve. Improvement is now something generation checks rather than something you have to remember. The workspace setting decides when it runs, never whether. Your OnePager and the transcript you read always agree, the improvement stays revertible, and regenerating a OnePager does not pay for the same work twice.
🐛 Notable fixes#
3 fixes
- The Generate OnePager button no longer goes missing on a Mac routed to the local engine. Two machines with identical settings could disagree about whether the button existed at all.
- Smart Actions no longer freeze the window while the model wakes up on a cold start.
- The cold start tells you which half it is in, loading the model or getting it ready, instead of going quiet for twenty seconds during the longer one.
Happy meetings! 🎙️