[Feature Request + Suggestions] Transition from documents / markdown to cms / relational db

[Rykarix]

Currently:
Everything works by installing bmad into a project directory. Then, as you progress through a task, everything works by writing and/or updating markdown documents that the bmad framework has to keep track of.

This causes a few problems, as I see it.

1. A lot of overhead in the context window that the ai has to keep track of > leads to many context related issues like rot, poisoning, distraction, etc
2. A lot of doc spamming in the project folder which causes a lot of problems managing different integrations (claude code vs cursor vs vscode vs blabla). It's also somewhat unsightly and means we have to mange multiple .ignore files etc. Not to mention all the other potential issues with some integrations that do semantic searching / indexing etc.

I propose the following:

1. Use Notion and utilise the Notion MCP to manage Project specific instructions, workflows & databases.
2. Use relational databases to provide inherent meaning to each workflow

Benefits of Using Notion:

Re: mcp related stuff: You can use notions mcp or make and customize your own.

Then you got the database stuff. Notion databases provide structured, queryable storage with typed properties (relations, rollups, formulas, multi-select). Workflow definitions and task states become database entries with explicit relationships rather than flat markdown files. Plus there's the whole 'multiple sourced' aspect of databases which is pretty cool.

Lastly it solves the issue of doc spamming project folders and creates this kind of centralised memory hub for documentation and workflows which live outside the project folder. No .ignore file management, no interference with semantic indexing, no conflicts between IDE integrations. Plus you have much more potential to expand and add functionality using notion as a kind of 'cms' for bmad.

Benefits of Using Neo4j:

Nodes support arbitrary properties including native temporal types. Each workflow step can carry created_at, modified_at, filepath, status flags, and navigation hints—keeping the node's content dedicated to actual contextual information while operational metadata lives in properties.

You can model "previous/next" as typed relationships (-[:NEXT_STEP]->, -[:DEPENDS_ON]->) rather than string metadata. Traversal queries become declarative graph operations rather than parsing exercises.

This then creates a kind of "separation of concerns" by allowing graph structures to handle workflow topology. Relationship properties handle conditions and priorities. Node properties handle timestamps and file references. Node content remains uncontaminated by infrastructure—containing only the payload the AI agent needs.

Query Efficiency: Relationship traversal is O(1) per hop. Finding "all pending steps after X" or "path from current to completion" doesn't require scanning tables or parsing document trees.

Neo4j also has an mcp

Thoughts?