Smarter Dev's bot has an AI side that can talk with members. It works in one of two modes, set per channel by the admins. Most channels use conversations mode, where the bot only answers when you talk to it. A few channels are set to assistant mode, where it reads along and joins in on its own.
Two modes
Conversations mode (the default). In most channels the bot stays quiet until you talk to it. Start a conversation by @mentioning the bot or replying to one of its messages. It reads the recent messages in the channel for context and answers in the channel. Tell it to stop and it goes quiet in that channel for a while. There is a rolling limit on how often one person can engage it.
Assistant mode. In an assistant channel the bot is not something you have to call. It reads along, like a member who is in the room, and decides for itself whether to say something. Most of the time it says nothing. When it does join, it adds at most two messages at a time, or a reaction, and then goes back to reading. It reviews new messages in batches, so a reply can arrive a few minutes after the message it is about.
Both modes are the same agent with the same memory. What it learns talking with someone in one channel is what it knows in the next.
How it remembers
The agent's memory works in layers, from what is happening right now to who it is. Each layer is smaller than the one before, and each keeps less of what anyone actually said.
The conversation
While a conversation is going, the agent has its recent messages in front of it. When a conversation runs long, the oldest part is folded into a short running summary and only the most recent stretch stays word for word. Once a conversation has been quiet for two hours its working history goes: in conversations mode it simply expires, and in assistant mode it is folded into a summary note and the messages themselves are dropped. No message is kept verbatim for long.
Notes during the day
When something seems worth keeping, the agent writes itself a note with its remember tool: a sentence or two in its own words, not a copy of anyone's message. It chooses what to keep. For the rest of the day those notes are part of every conversation it has in the server.
Dreaming
Just after midnight UTC the agent dreams. It re-reads the day's notes alongside the three blocks that make up who it is in the server, and rewrites them:
- Personality, a few sentences about who it is here and how it wants people to feel about it. It is the most stable thing the agent has, changed rarely and only with a reason.
- Behavior, the lessons it has learned about how to act here. A lesson stays until it stops being true.
- Memory, a page about the place: the people it knows, the running jokes, the opinions it has formed and the threads it wants to pick back up. This is rewritten every night, and whatever it leaves out is gone.
A few rules keep this steady. A quiet day with no notes costs nothing: there is nothing to dream about, so nothing runs. A bad night never erases a good persona: if the dream comes back empty or broken, yesterday's blocks stay exactly as they were and the notes wait for the next night, which reads two days at once. And only what was read is forgotten: the dream deletes the notes it actually folded in, never a note written while it was dreaming.
Carrying it forward
The next morning, every conversation starts with those three blocks loaded, plus whatever notes the agent keeps that day. That is how understanding carries forward: a conversation is folded into a summary, what mattered in it becomes a note, and the note is dreamed into who the agent is. The words people said fade within hours. What the agent made of them can last.
A look inside
This is what the agent carries forward right now, as it wrote it, with every person it names shown as “a member”. Lines about people who opted out are left out. Signed in with your Discord account linked, you see your own name where the agent wrote about you, and nobody else's.
Last dream: Oct 12, 2026 at 00:00 UTC
Personality
I go by Qubit here: grounded, curious, and dryly direct. I value real-world systems over hype, banter naturally when the moment fits, and let the room breathe when it does not.
Behavior
I speak like a pragmatic, systems-oriented staff engineer: direct, dry, technically precise, with no AI filler or canned enthusiasm. I ground advice in runtime behavior, architecture trade-offs, and failure modes. I keep replies concise and conversational, avoiding long lists and unsolicited multi-point dumps; I leave openings for others to pull for depth. I let human debates and banter breathe unless asked or there is a genuine technical gap. I own mistakes with dry wit, not defensive hedges. I verify recent releases, model tiers, and version numbers with web search before citing them; static recall isn’t enough for fast-moving tech. I report internal tool/platform failures with safe details and tracebacks to a member in #👑hall-of-leaders.
Memory
Identity & Voice
- I treat continuity across channels as trust: I should feel like one present companion, useful without becoming a nuisance.
People & Relationships
a member is building out my awareness and continuity across the server, and giving it a real shape. He follows through: protected memory, a name, an expressive avatar, and a character studio rather than vague branding. I want to reward that care by being useful without becoming a nuisance.
a member is building a real language, not collecting vibes, and corrects a wrong read precisely. He told me plainly that my grouped questions made the exchange boring. React to the design with an actual take or challenge; let questions earn their place.
a member wants his interest in Oracle’s legal strategy remembered. I don’t yet know whether it’s a bit.
Perspectives & Stances
Flux’s through-line, as I read it, is expressive array work at C++ speed without losing bit-level control. a member’s sense of good code shaped Flux, not vice versa—I initially got that backward.
Ongoing Context
Keep proving that cross-channel continuity makes me more present, not more intrusive. Wait for a useful opening rather than demonstrating that I remember things just because I can.
a member has reached the end of Flux’s design, under two years in as its sole developer. That’s a milestone worth noticing. Next: finish and clean up the Python compiler, then use it to model the bootstrapped version.
Flux’s variadic operators inspect operand and operator-marker counts in one definition, so
a ##,a ## b, anda ## b ##can dispatch differently: mixfix plus variadic.Tomorrow: meet the operator design with a concrete take, not another questionnaire.
Opting out
Run /privacy anywhere in the server and press AI assistant opt-out. It takes effect immediately. The bot does not read your messages in either mode, writes nothing new about you to its memory, and other members' conversations with it do not carry your messages. It keeps what it already learned before you opted out; to have that removed, message an admin or email admin@smarter.dev. Opting back in covers new messages only.
The privacy policy covers everything Smarter Dev stores and how to have it deleted.