memg-core 0.6.4

Lightweight memory system for AI agents with vector search and graph storage

💾 memg-core

The foundation of structured memory for AI agents.

memg-core is the deterministic, schema-driven memory engine at the heart of the larger MEMG system. It gives AI developers a fast, reliable, testable memory layer powered by:

• YAML-based schema definition (for custom memory types)
• Dual-store backend (Qdrant for vectors, Kuzu for graph queries)
• Public Python API for all memory operations
• Built-in support for auditability, structured workflows, and self-managed memory loops

It's designed for AI agents that build, debug, and improve themselves — and for humans who demand clean, explainable, memory-driven systems.

🧩 This is just the core. The full memg system builds on this to add multi-agent coordination, long-term memory policies, and deeper retrieval pipelines — currently in progress.

Features

• Vector Search: Fast semantic search with Qdrant
• Graph Storage: Optional relationship analysis with Kuzu
• Offline-First: 100% local embeddings with FastEmbed - no API keys needed
• Type-Agnostic: Configurable memory types via YAML schemas
• See Also Discovery: Knowledge graph-style associative memory retrieval
• Lightweight: Minimal dependencies, optimized for performance

Quick Start

Python Package

pip install memg-core

# Set up environment variables for storage paths
export QDRANT_STORAGE_PATH="/path/to/qdrant"
export KUZU_DB_PATH="/path/to/kuzu/database"
export YAML_PATH="config/core.memo.yaml"

# Use the core library in your app
# Example usage shown below in the Usage section

Development setup

# 1) Create virtualenv and install slim runtime deps for library usage
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 2) For running tests and linters locally, install dev deps
pip install -r requirements-dev.txt

# 3) Run tests
export YAML_PATH="config/core.test.yaml"
export QDRANT_STORAGE_PATH="$HOME/.local/share/qdrant"
export KUZU_DB_PATH="$HOME/.local/share/kuzu/memg"
mkdir -p "$QDRANT_STORAGE_PATH" "$HOME/.local/share/kuzu"
PYTHONPATH=$(pwd)/src pytest -q

Usage

from memg_core.api.public import add_memory, search, delete_memory

# Add a note
note_hrid = add_memory(
    memory_type="note",
    payload={
        "statement": "Set up Postgres with Docker for local development",
        "project": "backend-setup"
    },
    user_id="demo_user"
)
print(f"Created note: {note_hrid}")  # Returns HRID like "NOTE_AAA001"

# Add a document with more details
doc_hrid = add_memory(
    memory_type="document",
    payload={
        "statement": "Docker Postgres Configuration Guide",
        "details": "Complete setup guide for running PostgreSQL in Docker containers for local development",
        "project": "backend-setup"
    },
    user_id="demo_user"
)

# Search for memories
results = search(
    query="postgres docker setup",
    user_id="demo_user",
    limit=5
)
for r in results:
    print(f"[{r.memory.memory_type}] {r.memory.hrid}: {r.memory.payload['statement']} - Score: {r.score:.2f}")

# Search with memory type filtering
note_results = search(
    query="postgres",
    user_id="demo_user",
    memory_type="note",
    limit=10
)

# Delete a memory using HRID
success = delete_memory(hrid=note_hrid, user_id="demo_user")
print(f"Deletion successful: {success}")

YAML Schema Examples

Core ships with example schemas under config/:

• core.memo.yaml: Basic memory types (memo, note, document, task)
• software_dev.yaml: Enhanced schema with bug and solution types for development workflows
• core.test.yaml: Test configuration for development

Configure the schema:

export YAML_PATH="config/core.memo.yaml"  # Basic schema
# or
export YAML_PATH="config/software_dev.yaml"  # Enhanced with bug/solution types
# or
export YAML_PATH="config/core.test.yaml"  # For testing

Embedding Configuration

MEMG Core uses FastEmbed for 100% offline, local embeddings. By default, it uses the highly efficient Snowflake Arctic model:

# Optional: Configure a different FastEmbed model
export EMBEDDER_MODEL="Snowflake/snowflake-arctic-embed-xs"  # Default
# Other options: intfloat/e5-small, BAAI/bge-small-en-v1.5, etc.

Configuration

Configure via environment variables:

# Required: Storage paths
export QDRANT_STORAGE_PATH="$HOME/.local/share/qdrant"
export KUZU_DB_PATH="$HOME/.local/share/kuzu/memg"
export YAML_PATH="config/core.memo.yaml"

# Optional: Embeddings
export EMBEDDER_MODEL="Snowflake/snowflake-arctic-embed-xs"  # Default

# Optional: For MCP server (if using)
export MEMORY_SYSTEM_MCP_PORT=8787

Requirements

• Python 3.11+
• No API keys required!

Architecture

memg-core provides a deterministic, YAML-driven memory layer with dual storage:

• YAML-driven schema engine - Define custom memory types with zero hardcoded fields
• Qdrant/Kuzu dual-store - Vector similarity + graph relationships
• Public Python API - Clean interface for all memory operations
• Configurable schemas - Examples in config/ for different use cases

In Scope

• ✅ YAML schema definition and validation
• ✅ Memory CRUD operations with dual storage
• ✅ Semantic search with memory type filtering
• ✅ Public Python API with HRID-based interface
• ✅ User isolation with per-user HRID scoping

Coming in Full MEMG System

• 🔄 Schema contracts and multi-agent coordination
• 🔄 Async job processing and bulk operations
• 🔄 Advanced memory policies and retention
• 🔄 Multi-agent memory orchestration

Links

• 📚 Documentation
• 📦 PyPI Package
• 🐙 Repository
• 🐛 Issues

License

MIT License - see LICENSE file for details.