AI-Memory

Stop burning API tokens on file-by-file exploration. AI-Memory builds a self-updating knowledge graph of your entire codebase so your AI assistant answers instantly — with full context — instead of asking "can you show me that file?" repeatedly.

Before: You ask AI a question → AI opens 5 files → 8,000 tokens → still confused
After:  AI-Memory knows your code → AI queries the graph  → 500 tokens  → precise answer

The Problem (You Know This Pain)

You ask your AI assistant: "How do I add pagination to the user list?"

With AI-Memory: AI runs one internal query and gets list_users(self, db, skip=0, limit=100) with full body, types, and callers. 500 tokens. One shot. Correct answer.

How It Works

Supported Languages

DB Schemas: Table, index, column extraction from SQL and migration files.

Supported IDEs

After ai-memory init, your AI already knows about the graph. Zero IDE configuration.

Get Started (3 Commands)

# 1. Install
cd your-project
pip install ai-memory

# 2. Initialize — config, DB, gitignore, IDE rule files
ai-memory init

# 3. Scan your codebase — one time
ai-memory scan

Optional:

# Enable semantic search (~2 min one-time)
ai-memory embed

# Auto-update on file changes (background daemon)
ai-memory watch

What Happens After Init

your-project/
├── .ai-memory/           # Graph database (gitignored automatically)
│   └── graph.db          # SQLite: ~380 symbols, ~1850 relations, 34 deps
├── .ai-memory.toml       # Your project config
├── .windsurfrules        # Windsurf: "Use ai-memory for context"
├── .cursorrules          # Cursor: "Use ai-memory for context"
└── .gitignore            # Auto-added: .ai-memory/

CLI Commands

Core Workflow

Symbol Context

Search & Discovery

Analysis

Dependencies & Coverage

Export

MCP Server (Native IDE Integration)

For Claude Desktop, Windsurf, Cursor:

ai-memory-mcp --root /path/to/project

20 exposed tools: ai_memory_scan, ai_memory_stats, ai_memory_context, ai_memory_ask, ai_memory_semantic, ai_memory_flows, ai_memory_communities, ai_memory_schemas, ai_memory_find, ai_memory_dead_code, ai_memory_callers, ai_memory_todos, ai_memory_test_map, ai_memory_graph, ai_memory_cycles, ai_memory_inherits, ai_memory_api, ai_memory_search, ai_memory_coverage, ai_memory_deps.

Configuration

Edit .ai-memory.toml to customize:

scan_dirs = ["src", "app", "lib"]
extensions = [".py", ".js", ".ts", ".go", ".rs"]
ignore_patterns = ["node_modules", "venv", "__pycache__"]
output_format = "markdown"
max_context_tokens = 4000

Architecture

Source Code
    │
    ▼
Tree-sitter Parsers (Python, JS/TS, Go, Rust, Java, C#, C/C++, Ruby)
    │
    ▼
Symbol Extractor ──→ Symbols (functions, classes, imports)
│                     Body Extractor ──→ Full function bodies
│                     Type Extractor ──→ Type annotations
│
Relation Extractor ──→ Calls, Inherits, Imports
    │
    ▼
Import Resolver ──→ Cross-file links resolved
    │
    ▼
SQLite Graph DB ──→ nodes, edges, files, schemas, flows, communities
    │
    ▼
┌─────────────────┬─────────────────┬─────────────────┐
│  Semantic       │  Execution      │  Community      │
│  Embeddings     │  Flows          │  Detection      │
│  (MiniLM)       │  (BFS from      │  (Leiden        │
│                 │   entry points) │   algorithm)    │
└─────────────────┴─────────────────┴─────────────────┘
    │
    ▼
CLI / MCP Server / Watch Daemon
    │
    ▼
AI Assistant (Windsurf, Cursor, Claude Desktop)