This is the devlog about this project, here some details:
Experiment window: 7 days
Test environment: Discord server Snug Nook (snugnook.org)
This was a 7-day experiment to see if a single main.ts process could feel like a persistent person instead of a command bot. The goal was technical: stable behavior, memory over time, and believable pacing in chat.
Core setup:
Adrian runs on a simple stack:
-
Deno runtime
-
discord.js-selfbot-v13 for Discord events and presence
-
Local MLX endpoint (
http://127.0.0.1:8080/v1/chat/completions) for generation
-
SQLite for user memory (
users, user_facts)
Everything is wired together inside main.ts.
How it behaved:
The flow is straightforward:
- Listen to incoming messages and filter obvious noise.
- Score whether a message is worth replying to (ping, DM, question, reply context, etc.).
- Buffer user input for ~5 seconds so split messages can be merged.
- Generate a response with persona constraints.
- Clean and validate output before sending.
On top of replies, Adrian also:
- Updates mood/presence every 10 minutes from recent context.
- Posts occasional autonomous “icebreaker” thoughts.
- Reacts to hype/meme messages and sometimes joins reaction bandwagons.
Memory and state:
Two memory layers were used:
-
Persistent (SQLite): trust score, vibe summary, extracted user facts.
-
In-memory: short rolling context, pending reply buffers, recent ping history.
This gave decent continuity across chats while keeping runtime logic lightweight.
Guardrails that helped:
A few practical safeguards carried most of the stability:
- Coherence checks to reject weird/noisy outputs.
- Strict formatting prompts for mood, vibe, and fact extraction.
- Fallbacks when generation failed or came back too long.
- Sleep/wake scheduling so behavior had natural downtime.
Limits we hit:
The design worked, but it has clear limits:
- Persona consistency is mostly prompt-based.
- One global
isProcessing lock can bottleneck multiple concurrent conversations.
- Some conversational state is process-local and resets on restart.
- If the local LLM server is down, behavior degrades hard.
Final takeaway (7 days in Snug Nook):
In Snug Nook (snugnook.org), this looked less like a classic chatbot and more like a small autonomous loop: ingest events, reason with local inference, store lightweight memory, and keep running. It stayed technically coherent for the full 7-day window, with most failures handled by simple guardrails rather than heavy infrastructure.
ps: I’ve attached some of the funniest moments of his while talking with fellow discord members
