Living Town

Project intro

Living Town is a graph-memory engine for multi-agent systems, presented as a town that gossips. Ten NPC agents live in Thornbury, and each one's memory is a neighbourhood in a FalkorDB graph: what they saw (episodic), what they believe and who told them (semantic, with confidence, stance and hop count), and what they know how to do (procedural). Steal the blacksmith's silver hammer and only the blacksmith sees it; over the next simulated hours the story travels along social edges — Bram tells Mira at the tavern, Mira tells Captain Oswin, Oswin posts an investigation, the deputy claims it because the database only lets an agent claim a task it KNOWS_HOW to do — losing confidence at every hop (speaker confidence x listener trust x 0.9, doubted below 0.3) until the guard who saw nothing confronts you at dawn. Ask any NPC 'How do you know?' and the answer is not generated text but the exact provenance path the query returned, drawn as a red thread across the map; ask Sister Agnes and she says nothing in her memory connects to it, because no path means no answer, not a guess. Every player gets an isolated graph (town:player), so two players side by side live in two different worlds, and returning under the same name resumes the town's memory of you. The same 19 memory tools are exposed over MCP, so any MCP client can open a town and ask who told whom. Without an API key the NPCs run on rule-based agents; with an Anthropic key, dialogue is generated by Claude over the same fixed memory catalogue, with any mention the graph never returned thrown away. Built for Graph Hacks: Context for AI Agents (WeMakeDevs x FalkorDB). Free-tier limits, as the author states: the Render app can sleep so the first load may take a minute, and a paused free FalkorDB Cloud instance would break the live site (the same code runs locally as fallback).