A local knowledge graph for my game-design ideas. Concepts, mechanics and monetization patterns end up as connected nodes in a graph database, become semantically searchable, and Claude reads, writes and links them directly through an MCP server.
When I think about games, ideas pile up constantly: a mechanic from a screenshot, a monetization trick from a YouTube video, a loop from some roguelike. Normally those end up in a note and are gone three weeks later, or at least impossible to find again. I wanted a memory for them that doesn't just store ideas but actually connects them — and where I can check whether I've had something like it before, without knowing the exact term.
Every concept becomes a node in the graph and gets a meaning fingerprint of 768 numbers from a local embedding model. That lets me find related ideas later through semantic search even when they're worded completely differently — the question "do we already have something with stages and gates?" also hits a concept I called "level progression". I link the nodes with typed edges like combines with, is variant of, monetizes or inspired by, so clicking through surfaces new combinations I hadn't thought of before.
The graph is fed by a handful of small Claude skills: one pulls concepts out of screenshots (shops, skill trees, HUDs), one out of gameplay videos and trailers, and a dedicated trend scout automatically dumps viral Roblox memes and clonable Steam loops into it. Another skill just asks — "is there already a concept for X?" — before I add something twice. Claude talks to the database directly through an MCP server and translates my instructions into graph and vector queries.
Everything except Claude runs locally on my homelab: Postgres with Apache AGE holds the graph, pgvector stores the embeddings and computes similarities, Ollama generates the embeddings offline. A read-only visualization shows the whole graph in the browser, edges colored by relation type — by now that's 462 concepts and 887 relations. The knowledge graph stays entirely with me.
Tech-Stack: PostgreSQL 16, Apache AGE, pgvector, Ollama (nomic-embed-text), MCP server (TypeScript), Docker