enyal 0.1.2

Persistent, queryable memory for AI coding agents

Enyal

Persistent, queryable memory for AI coding agents.

Enyal gives AI agents like Claude Code durable context that survives session restarts. Every conversation becomes accumulated institutional knowledge—facts, preferences, decisions, and conventions that persist and grow.

Features

• Persistent Memory: Context survives restarts, crashes, and process termination
• Semantic Search: Find relevant context using natural language queries (384-dim embeddings via all-MiniLM-L6-v2)
• Hierarchical Scoping: Global → workspace → project → file context inheritance
• Fully Offline: Zero network calls during operation
• Cross-Platform: macOS (Intel + Apple Silicon), Linux, and Windows
• MCP Compatible: Works with Claude Code, Cursor, Windsurf, Kiro, and any MCP client

Quick Start

Get up and running in under 2 minutes:

1. Install

# Using uvx (recommended - no installation needed)
uvx enyal serve

# Or install with pip
pip install enyal

2. Configure Your MCP Client

Universal configuration (works with Claude Code, Cursor, Windsurf, Kiro):

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["enyal", "serve"]
    }
  }
}

For macOS Intel users (requires Python 3.11 or 3.12):

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["--python", "3.12", "enyal", "serve"]
    }
  }
}

3. Start Using

You: Remember that this project uses pytest for all testing
Assistant: [calls enyal_remember] Stored context about testing framework

You: What testing framework should I use?
Assistant: [calls enyal_recall] Based on stored context, this project uses pytest.

Platform Support

Platform	Python 3.11	Python 3.12	Python 3.13
macOS Apple Silicon	uvx enyal serve	uvx enyal serve	uvx enyal serve
macOS Intel	uvx --python 3.11 enyal serve	uvx --python 3.12 enyal serve	Not supported*
Linux	uvx enyal serve	uvx enyal serve	uvx enyal serve
Windows	uvx enyal serve	uvx enyal serve	uvx enyal serve

*macOS Intel + Python 3.13 is not supported due to PyTorch ecosystem constraints.

Installation Methods

Method 1: uvx (Recommended for MCP)

# Most platforms (auto-selects Python)
uvx enyal serve

# macOS Intel (explicit Python version)
uvx --python 3.12 enyal serve

# With model preloading for faster first query
uvx enyal serve --preload

Method 2: pipx

# Install globally
pipx install enyal

# Run server
enyal serve

Method 3: pip

# Using uv (recommended)
uv add enyal

# Using pip
pip install enyal

# Run server
enyal serve

MCP Integration

Enyal works with any MCP-compatible client. The configuration is the same across platforms—only the command may vary for macOS Intel.

Claude Code

File locations:

• Project: .mcp.json (in project root)
• User: ~/.claude/.mcp.json

Standard configuration:

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["enyal", "serve"],
      "env": {
        "ENYAL_DB_PATH": "~/.enyal/context.db"
      }
    }
  }
}

macOS Intel configuration:

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["--python", "3.12", "enyal", "serve"],
      "env": {
        "ENYAL_DB_PATH": "~/.enyal/context.db"
      }
    }
  }
}

CLI setup:

# Standard
claude mcp add-json enyal '{"command":"uvx","args":["enyal","serve"]}'

# macOS Intel
claude mcp add-json enyal '{"command":"uvx","args":["--python","3.12","enyal","serve"]}'

Claude Desktop

File locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Configuration:

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["enyal", "serve"],
      "env": {
        "ENYAL_DB_PATH": "~/.enyal/context.db"
      }
    }
  }
}

Cursor

File locations:

• Global: ~/.cursor/mcp.json
• Project: .cursor/mcp.json

Configuration:

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["enyal", "serve"],
      "env": {
        "ENYAL_DB_PATH": "~/.enyal/context.db"
      }
    }
  }
}

UI setup: File → Preferences → Cursor Settings → MCP

Windsurf

File location: ~/.codeium/windsurf/mcp_config.json

Configuration:

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["enyal", "serve"],
      "env": {
        "ENYAL_DB_PATH": "~/.enyal/context.db"
      }
    }
  }
}

UI setup: Windsurf Settings → Cascade → MCP, or use the Plugin Store

Kiro

File locations:

• Global: ~/.kiro/settings/mcp.json
• Project: .kiro/settings/mcp.json

Configuration:

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["enyal", "serve"],
      "env": {
        "ENYAL_DB_PATH": "~/.enyal/context.db"
      },
      "autoApprove": ["enyal_recall", "enyal_stats", "enyal_get"]
    }
  }
}

UI setup: Click the Kiro ghost tab → MCP Servers → "+"

See docs/INTEGRATIONS.md for detailed platform-specific guides.

Available Tools

Tool	Description
enyal_remember	Store new context with metadata (facts, preferences, decisions, conventions, patterns)
enyal_recall	Semantic search for relevant context with filtering by scope and type
enyal_forget	Remove or deprecate context (soft-delete by default, hard-delete optional)
enyal_update	Update existing entries (content, confidence, tags)
enyal_get	Retrieve a specific entry by ID with full metadata
enyal_stats	Get usage statistics and health metrics

Content Types

Type	Use For	Example
fact	Objective information	"The database uses PostgreSQL 15"
preference	User/team preferences	"Prefer tabs over spaces"
decision	Recorded decisions	"Chose React over Vue for frontend"
convention	Coding standards	"All API endpoints follow REST naming"
pattern	Code patterns	"Error handling uses Result<T, E> pattern"

Scope Levels

Scope	Applies To	Example Path
global	All projects	(none)
workspace	Directory of projects	/Users/dev/projects
project	Single project	/Users/dev/myproject
file	Specific file	/Users/dev/myproject/src/auth.py

CLI Usage

Enyal provides a command-line interface for direct interaction:

# Store context
enyal remember "Always use pytest for testing" --type convention --scope project

# Search context
enyal recall "testing framework" --limit 5

# Get entry details
enyal get <entry-id>

# View statistics
enyal stats

# Remove context
enyal forget <entry-id>

# Run MCP server
enyal serve --preload

Options:

• --db PATH — Custom database path
• --json — Output in JSON format

See docs/CLI.md for complete CLI reference.

Python Library

from enyal.core.store import ContextStore
from enyal.core.retrieval import RetrievalEngine
from enyal.models.context import ContextType, ScopeLevel

# Initialize store
store = ContextStore("~/.enyal/context.db")
retrieval = RetrievalEngine(store)

# Remember something
entry_id = store.remember(
    content="Always use pytest for testing in this project",
    content_type=ContextType.CONVENTION,
    scope_level=ScopeLevel.PROJECT,
    scope_path="/Users/dev/myproject",
    tags=["testing", "pytest"]
)

# Recall relevant context
results = retrieval.search(
    query="how should I write tests?",
    limit=5,
    min_confidence=0.5
)

for result in results:
    print(f"{result.score:.2f}: {result.entry.content}")

# Update context
store.update(entry_id, confidence=0.9, tags=["testing", "pytest", "unit-tests"])

# Get specific entry
entry = store.get(entry_id)

# Get statistics
stats = store.stats()
print(f"Total entries: {stats.total_entries}")

Configuration

Environment Variables

Variable	Default	Description
ENYAL_DB_PATH	~/.enyal/context.db	Database file location
ENYAL_PRELOAD_MODEL	false	Pre-load embedding model at startup
ENYAL_LOG_LEVEL	INFO	Logging level (DEBUG, INFO, WARNING, ERROR)

Database Location

The default database is stored at ~/.enyal/context.db. This single SQLite file contains:

• All context entries and metadata
• Vector embeddings for semantic search
• Full-text search index

Troubleshooting

Installation Fails on macOS Intel

Symptom: Error about torch/PyTorch wheels not found

Cause: PyTorch doesn't provide wheels for macOS Intel + Python 3.13

Solution: Use Python 3.11 or 3.12:

uvx --python 3.12 enyal serve

MCP Server Not Connecting

1. Check uvx is installed:

   uvx --version
   

2. Test server manually:

   uvx enyal serve
   # Should start without errors, waiting for MCP protocol
   

3. Enable debug logging:

   {
     "mcpServers": {
       "enyal": {
         "command": "uvx",
         "args": ["enyal", "serve", "--log-level", "DEBUG"]
       }
     }
   }
   

4. Check server status:

   • Claude Code: /mcp command
   • Cursor: Settings → MCP → check status
   • Windsurf: Cascade → Plugins
   • Kiro: Ghost tab → MCP Servers

Slow First Query

The first query loads the embedding model (~80MB). This takes ~1-2 seconds. Subsequent queries are fast (~34ms).

To pre-load the model at startup:

{
  "mcpServers": {
    "enyal": {
      "command": "uvx",
      "args": ["enyal", "serve", "--preload"]
    }
  }
}

Database Locked Error

If you see "database is locked" errors, ensure only one MCP server instance is running per database file. Use different ENYAL_DB_PATH values for different projects if needed.

Permission Errors

On macOS/Linux, ensure the database directory exists and is writable:

mkdir -p ~/.enyal
chmod 755 ~/.enyal

Architecture

Enyal uses a unified SQLite database with:

• Relational storage for metadata and attributes
• sqlite-vec for vector similarity search (384-dim embeddings)
• FTS5 for keyword search
• WAL mode for concurrent access

See docs/ARCHITECTURE.md for detailed design decisions.

Development

# Clone repository
git clone https://github.com/seancorkum/enyal.git
cd enyal

# Install with dev dependencies
uv sync --all-extras

# Run tests
uv run pytest

# Type checking
uv run mypy src/enyal

# Linting
uv run ruff check src/enyal

Performance

Benchmarked on Intel Mac with Python 3.12:

Metric	Target (p95)	Measured (p95)	Status
Cold start (model load + first query)	<2000ms	~1500ms	✓
Warm query latency	<50ms	~34ms	✓
Write latency	<50ms	~34ms	✓
Concurrent reads (4 threads)	<150ms	~85ms	✓
Memory (100k entries estimated)	<500MB	~35MB	✓

Embedding model: all-MiniLM-L6-v2 (22M params, 384 dimensions)

Run benchmarks:

uv run python benchmarks/benchmark_performance.py

License

MIT