Abstract
Raw prose becomes useful agent memory only after we model identity, relationships, scope, and evidence—not merely embeddings. DungeonBuddy turns years of campaign prose into a world the GM can inspect and an agent can query without pretending to know things the graph cannot support.
TTRPG campaigns generate years of prose but very little durable structure. DungeonBuddy is an experiment in turning session recaps, prep notes, NPC files, and worldbuilding documents into a provenance-aware World Supergraph.
The system extracts typed objects and relationships, resolves identity, preserves source evidence, and materializes revisioned semantic objects. Those objects can then be projected into different workflows—session planning, live play, graph review, or a Hermes agent—without each surface inventing its own version of the world.
This talk follows one campaign object from raw Markdown through graph construction, contextual projection, agent retrieval, and a governed contribution path. The central challenge is not merely helping an agent find an answer. It is allowing humans and agents to interact with a shared world model while remaining explicit about who may change it, what counts as evidence, and when a proposed change becomes durable campaign memory.
Date
7/23/26
This will be my first tech talk
Code of Conduct
Abstract
Raw prose becomes useful agent memory only after we model identity, relationships, scope, and evidence—not merely embeddings. DungeonBuddy turns years of campaign prose into a world the GM can inspect and an agent can query without pretending to know things the graph cannot support.
TTRPG campaigns generate years of prose but very little durable structure. DungeonBuddy is an experiment in turning session recaps, prep notes, NPC files, and worldbuilding documents into a provenance-aware World Supergraph.
The system extracts typed objects and relationships, resolves identity, preserves source evidence, and materializes revisioned semantic objects. Those objects can then be projected into different workflows—session planning, live play, graph review, or a Hermes agent—without each surface inventing its own version of the world.
This talk follows one campaign object from raw Markdown through graph construction, contextual projection, agent retrieval, and a governed contribution path. The central challenge is not merely helping an agent find an answer. It is allowing humans and agents to interact with a shared world model while remaining explicit about who may change it, what counts as evidence, and when a proposed change becomes durable campaign memory.
Date
7/23/26
This will be my first tech talk
Code of Conduct