DEV Community

Milenko Mitrovic
Milenko Mitrovic

Posted on

Stop Re-explaining Your Codebase to AI — Give It Permanent Memory Instead

#ai #opensource #python #devtools

I Built a Local Knowledge Graph That Tells AI Agents "What Breaks If I Change This?"

Every time I start a new AI coding session, the same thing happens: I re-explain my architecture, my conventions, my service dependencies. The AI doesn't remember. It doesn't know that store.py is imported by 8 other modules, that Milenko owns the auth service, or that postgres is a single point of failure.

So I built Knowledge Master — a local knowledge graph that gives AI agents permanent memory of your codebase.

What makes it different from "just another RAG"

Most RAG tools embed your code into vectors and return "similar chunks." That's fancy grep. It doesn't understand relationships.

Knowledge Master builds an actual graph:

Repo → USES_TECH → Python, FastAPI, Docker
Repo → DEFINES_SERVICE → auth-service → DEPENDS_ON → postgres
Person → AUTHORED → Document → IMPORTS → Document
Function → DEFINED_IN → File → IN_REPO → Repo
Enter fullscreen mode Exit fullscreen mode

This lets it answer questions that flat search can't:

"What breaks if I change this?" 

$ km blast-radius store.py
💥 Blast radius: store.py
├── Definite impact
│   ├── 📄 cli.py (IMPORTS)
│   ├── 📄 server.py (IMPORTS)
│   ├── 📄 web.py (IMPORTS)
│   ├── 📄 api.py (IMPORTS)
│   └── 📄 connectors.py (IMPORTS)
├── Likely affected
│   └── ⚙️ falkordb (Service, owns affected file)
└── Possibly affected
    └── 👤 User
![ ](https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/mk0kohpdg09euf0ibazz.png) (Person, AUTHORED affected file)
Enter fullscreen mode Exit fullscreen mode

This uses real static analysis — Python AST, tree-sitter for TypeScript/Go/Rust/Java/C# — to trace actual import dependencies. Not text similarity.

"Is it safe to touch this?" 

$ km safe-to-change auth/service.py
Risk: RISKY
├── Blast radius: 12 entities
├── Test coverage: yes (3 test files)
└── Affected: api.py, gateway.py, user-service...
Enter fullscreen mode Exit fullscreen mode

Combines blast radius + test coverage detection into a risk score: safe / risky / dangerous.

"Who owns this code?" 

$ km who-owns src/payments/stripe.py
Owner: Alex (weight: 0.85)
Enter fullscreen mode Exit fullscreen mode

Git blame weighted by recency — recent changes count more than ancient history.

How it works 

Your repos
    ↓ km index
┌─────────────────────────────┐
│  FalkorDB (Graph + Vector)  │
│                             │
│  Nodes: Repo, Document,    │
│  Person, Service, Tech,    │
│  Function, Convention      │
│                             │
│  Edges: IMPORTS, DEPENDS_ON,│
│  AUTHORED, OWNS, USES_TECH │
│                             │
│  + Vector embeddings for    │
│    semantic search          │
└─────────────────────────────┘
    ↓ MCP protocol
Your AI agent (Claude, Cursor, Copilot, Kiro)
Enter fullscreen mode Exit fullscreen mode

Everything runs locally:

  • FalkorDB — graph database with built-in vector search (single Docker container)
  • Ollama — local embeddings (nomic-embed-text, 274MB)
  • Tree-sitter — structural code parsing for 7 languages
  • No cloud, no API keys, no data leaves your machine

7 Languages supported

Language Static Analysis
Python AST import graph, function/class extraction
TypeScript/JS tree-sitter imports, exports
Go tree-sitter imports, exported functions/types
Rust tree-sitter use/mod, pub items
Java tree-sitter imports, public classes/methods
C# tree-sitter using directives, public members
Terraform Module dependencies, resource/variable extraction

MCP Server — AI agents use it directly

Knowledge Master exposes 8 tools via the Model Context Protocol:

Tool What the AI agent can do
search Semantic search with re-ranking
blast_radius "What breaks if I change X?"
safe_to_change Risk assessment
who_owns File ownership
check_conventions "Does this follow our patterns?"
index_repo Add a new repo
index_directory Add docs
get_status Knowledge base stats

Setup for any AI tool is one command:

km setup cursor    # or claude, kiro, copilot, amazonq
Enter fullscreen mode Exit fullscreen mode

Try it 

pipx install knowledge-master
km start
km index ~/your-project
km search "how does authentication work"
km blast-radius auth/service.py
km safe-to-change auth/service.py
km serve  # web UI with graph visualization
Enter fullscreen mode Exit fullscreen mode

Links

I'd love feedback — especially on:

  • What languages/frameworks should be prioritized next?
  • Would you prefer deeper call-graph analysis or more connectors (Slack, Jira)?
  • Is the MCP integration the right bet, or would a VS Code extension be more useful?

MIT licensed, open source, built for developers who are tired of re-explaining their codebase to AI.

Top comments (0)