Murder Mystery
Agents in Context Toronto: a MongoDB HackathonMurder Mystery is a multiplayer detective game on Discord where AI suspects have memory you can govern. Teams play the same case at the same time, interrogating voiced suspects in their own words, presenting evidence to break their stories, following leads, and making one accusation to solve the case. Each team plays in its own copy of the world, so what one team uncovers changes only its own suspects. Ask the butler the same question from two teams and you'll hear the same voice tell two different stories. A murder mystery is a game about who knows what, and that's exactly where AI characters fail: they leak the killer, forget their alibi, or confess because you asked nicely. We treated each of those as a memory problem and solved it in MongoDB Atlas, not in the prompt. Every memory belongs to one team and lists which suspects know it, and every vector search is filtered by those rules, so a suspect can't reveal what it was never allowed to retrieve. Memories are never edited. When a witness recants, the old memory is superseded in a transaction with its history and an audit trail kept, and any lead it supported stops working immediately. Rumours expire on their own, and confessions are only possible once a suspect's story has actually been broken. Every reply is recorded as a trace of the memories behind it, and every action is checked against current memory before it runs, so you can see that changing a memory changed the suspect's behaviour. Built with MongoDB Atlas and Atlas Vector Search, Voyage AI embeddings, a Gemini-powered agent built for Atlas Agent Engine, discord.py, and ElevenLabs voices.




