memuri 0.1.6.dev25

Memuri is a self‑hosted memory infrastructure for AI applications, offering a unified, pluggable SDK to persist and retrieve conversational context with sub‑second latency. It enables any AI agent or assistant to remember key information across sessions—so your apps never forget what matters, while retaining full flexibility to swap storage backends, embedding models, and retrieval strategies.

Memuri - High-Performance Memory SDK

Memuri is a pip-installable SDK for high-performance, pluggable conversational memory services. Designed for local, self-hosted deployment with sub-second latency and production-grade practices.

Key Features

• Self-hosted: No external dependencies or licenses required
• High Performance: Sub-100ms memory operations with Faiss and Redis caching
• Pluggable Architecture: Swap vector databases, task queues, and LLMs without code changes
• Production-grade: Integrated OpenTelemetry traces, health checks, and robust CI/CD
• Memory Categories: Organize context by category (PERSONAL, TASK, QUESTION, etc.)
• Feedback Loop: Adaptive classification based on user feedback
• Reranking: Advanced cross-encoder reranking for more relevant results
• Flexible Configuration: Easy configuration with environment variables, dictionary-based config, or direct settings
• Multiple Embedding Providers: Support for OpenAI, Google Gemini, Azure, and Sentence Transformers

Installation

You can install Memuri directly from PyPI:

pip install memuri

Or using Poetry:

poetry add memuri

Getting Started

Basic Usage

from memuri import Memuri

# Initialize memory with default settings
memory = Memuri()

# Add a memory item
memory.add_memory(content="John's favorite color is blue", category="PERSONAL")

# Search for relevant memories
results = memory.search_memory("What does John like?")
print(results)

Using Config Dictionary

import os
from memuri import Memuri

# Set API key in environment variable
os.environ["OPENAI_API_KEY"] = "your_api_key"

# Create config with specific provider settings
config = {
    "embedder": {
        "provider": "openai",
        "config": {
            "model": "text-embedding-3-small"
        }
    }
}

# Initialize with config
memory = Memuri.from_config(config)

Storing Chat Conversations

# Add a conversation as memory
messages = [
    {"role": "user", "content": "I'm planning to watch a movie tonight. Any recommendations?"},
    {"role": "assistant", "content": "How about sci-fi? I recommend Interstellar."},
    {"role": "user", "content": "I love sci-fi movies!"}
]

# Store the conversation with user ID
await memory.add(messages, user_id="john")

Architecture

Memuri is designed around a layered memory system:

1. Short-Term Memory: HNSW In-Memory Index + Redis LRU Cache
2. Long-Term Memory: Pluggable Vector Stores (pgvector, Milvus, Qdrant, Redis Vector)
3. Memory Triggers: Category classifiers and rule engine for contextual decisions
4. Feedback Loop: Continuous adaptation based on user interactions

Performance Testing

Memuri includes a comprehensive latency testing framework to measure performance metrics:

Using the Latency Test Script

# Run all tests with default settings
cd src/memuri/tests
./run_latency_tests.py

# Run only add memory tests with 20 iterations
./run_latency_tests.py --test-type add --iterations 20

# Run only search tests with custom database URL
./run_latency_tests.py --test-type search --db-url postgresql://user:pass@localhost:5432/mydb

# Set log level to DEBUG for more detailed output
./run_latency_tests.py --log-level DEBUG

# Don't save results to a file
./run_latency_tests.py --no-save

Test Output

The test runner will output detailed statistics for each operation:

=== Add Memory Benchmark Results ===
Total operations: 10
Mean latency: 831.96ms
Median latency: 990.54ms
Min latency: 322.88ms
Max latency: 1361.35ms
95th percentile: 1273.73ms
Standard deviation: 387.65ms
==================================

=== Search Memory Benchmark Results ===
Total operations: 10
Mean latency: 817.86ms
Median latency: 804.20ms
Min latency: 410.45ms
Max latency: 1388.41ms
95th percentile: 1283.28ms
Standard deviation: 359.60ms
==================================

Programmatic Usage

You can also use the LatencyBenchmark class in your own code:

import asyncio
from memuri import Memuri
from memuri.tests.test_latency import LatencyBenchmark

async def measure_performance():
    client = Memuri()
    benchmark = LatencyBenchmark("Custom Test")
    
    # Measure add_memory operation
    for i in range(5):
        await benchmark.measure_operation(
            client.add_memory,
            content=f"Test memory {i}",
            category="FACT"
        )
    
    # Print stats
    benchmark.print_stats()
    
    # Get stats as dictionary
    stats = benchmark.get_stats()
    print(f"95th percentile latency: {stats['p95']:.2f}ms")

asyncio.run(measure_performance())

Development and Deployment

Publishing to PyPI

The project includes a script to handle publishing to PyPI with automatic version management:

# From project root
./scripts/upload_to_pypi.sh

This script will:

1. Automatically increment version numbers (patch, minor, or major)
2. Run tests before building
3. Build and upload the package to PyPI
4. Create git tags and commits for the release

Next Steps

• Check the Quick Start guide
• Learn about Configuration options
• Explore API Reference
• Learn advanced patterns with Cookbooks
• View Examples for complete solutions

Updates and Improvements

Latest Updates (v0.2.0)

We've enhanced Memuri with several key improvements:

1. Enhanced Category System

   • Hierarchical memory categorization with main categories and subcategories
   • 15 main categories and 45+ subcategories for precise memory organization
   • Backward compatibility with legacy categories

2. Improved OpenAI Embedding Support

   • Dynamic client configuration with proper error handling
   • Support for all OpenAI embedding models including text-embedding-3-small/large
   • Advanced configuration options (proxies, Azure integration, custom parameters)

3. Feedback System Enhancement

   • Track and analyze feedback for both categories and subcategories
   • Improved classifier training based on user feedback
   • Parent-child relationship awareness in memory classification

4. Configuration Flexibility

   • Enhanced from_config method for cleaner, more intuitive setup
   • Direct parameter passing to embedding and LLM services
   • Better factory patterns for service initialization

Check the documentation for details on how to use these new features.