MDEMG
A graph memory and reasoning substrate for AI agents: multi-dimensional, emergent, and reinforced by use.
- My role
- Sole developer
- Status
- In active development
- Since
- 2024
- Source
- github.com/reh3376/mdemg
- Go
- vector retrieval
- knowledge graphs
- RAG
- LLM re-ranking
The problem
An agent with a vector store can find text that resembles a question. It cannot easily say what relates to what, or why one association should matter more than another after a thousand uses.
The approach
A multi-dimensional emergent memory graph: vector retrieval and knowledge graph structure together, with agent memory and state management on top, LLM re-ranking over the candidates, and Hebbian reinforcement so associations that keep proving useful strengthen and the rest decay.
Outcomes
Two years of work and still active, with runners and clients across several platforms: a self-contained sidecar runner, a Linux installer, a macOS menu-bar app, a Windows build, an observability sidebar and a Neovim plugin.
The public repository dates from 2026; the work began in 2024.