m_flow

Why we included this project

Most RAG systems rank chunks by vector similarity, which is why they keep returning documents that look related but miss the actual answer. M-flow treats the knowledge graph as the scoring engine instead of an accessory: vector search finds entry points, then typed, semantically weighted edges propagate evidence and rank each unit by the strongest reasoning path back to the query. That shift matters for questions like 'why was Maria upset at Monday's standup?', where keyword overlap lands on a generic article about running standups rather than the real cause. Knowledge is organized into a layered cone graph of episodes, facets, and entities, and retrieval hands back a coherent episode bundle the LLM can reason from. It is a Python library with an MCP skill, so it slots into existing agent and retrieval pipelines where long-term memory is the goal.