Memory that survives the thread
Most assistants forget you the moment the thread ends. Nova treats every turn as something worth keeping.
Ask any chat assistant what you decided last Tuesday and you get a polite blank. The context window is its whole world, and the window closes with the thread. For a team running operations in chat, that is not a quirk. It is the reason nobody trusts the assistant with anything that matters.
Nova takes the opposite stance. Every turn becomes a compact observation: who said what, what was decided, what was done about it. Each observation is embedded and stored. Nothing waits for someone to type “remember this”.
What a turn becomes
Here is what happens after every exchange. Nova distils the turn into the part you would have written on a sticky note: who asked, what was agreed, what was done, what is still open. That observation is embedded and indexed for both full-text and vector search. It is not a transcript dump. Transcripts are where decisions go to get lost.
Two kinds of recall, fused
Keyword search finds the exact ticket number. Vector search finds the conversation you half remember. Each alone is wrong half the time. Nova runs both and fuses the rankings, so a question like “what did we agree about the deploy window” lands on the right exchange whether or not anyone used those words.
Say someone asks on Thursday: “did we ever decide what to do about the flaky staging deploy?” There is no ticket number in that sentence. Keyword search alone misses Monday’s thread, which was about a “rollout window”. Vector search alone might surface three vaguely similar incidents from the spring. Fused, the ranking lands on Monday, and Nova answers with the decision, who made it, and the step that was never finished.
“It remembers every decision we’ve made and ships its own fixes overnight.”
Memory per person, not per bot
In Telegram and Discord, Nova routes every person to their own persistent session. Your context is yours. A group chat gets its own, too. The agent answering the on-call engineer at 3am is carrying that engineer’s history, not a blended soup of everyone’s.
This matters more than it sounds. The ops lead and the finance lead ask about “the report” and mean different reports. A single shared memory would make Nova confidently wrong for one of them. Per-person sessions make it quietly right for both.
The window manages itself
Long conversations blow past any context limit. Nova compacts automatically, keeping the observations and dropping the noise, so a session that started in March still works in October without anyone restarting it. And because every observation stays searchable across all history, compaction never means amnesia. The detail is a query away.
What this changes
Trust compounds. In the first week, people ask Nova for the things they would ask a new colleague: summarise this, draft that. By the third week they are saying “do what we did last time”, and Nova knows what that means. That is the moment an assistant becomes an operator.
Memory is not a feature we bolted onto an assistant. It is the thing that makes an agent trustworthy enough to hand work to. Everything else on this site rests on it.