remina 0.1.5

Give your AI agents human-like memory. The pluggable memory layer for AI applications.

Remina

Memory infrastructure for AI applications.

Remina is the memory layer for AI systems. Extract facts, persist context, and retrieve relevant memories — with pluggable infrastructure that fits your stack.

from remina import Memory

memory = Memory()

# Extract and store facts
memory.add("I'm John, I work at Google as a software engineer.", user_id="john")

# Semantic retrieval
results = memory.search("What is John's profession?", user_id="john")
# → ["Works as a software engineer at Google"]

Why Remina

AI applications without persistent memory accumulate technical debt: lost context, repeated user friction, degraded personalization. Remina addresses this at the infrastructure layer.

Core capabilities:

• Automatic fact extraction — LLM-powered extraction from conversations
• Hybrid retrieval — Vector similarity + temporal decay + importance weighting
• Deduplication — Similarity-based duplicate prevention
• Pluggable architecture — Swap storage, vectors, embeddings, and LLM providers independently

Installation

pip install remina

Provider extras: openai, anthropic, gemini, pinecone, qdrant, chroma, postgres, mongodb, redis.

Quick Start

from remina import Memory

# Defaults: OpenAI embeddings + Chroma vectors + SQLite storage
memory = Memory()

# Extract facts from conversation
memory.add(
    messages=[
        {"role": "user", "content": "I just moved to Seattle"},
        {"role": "assistant", "content": "How do you like it?"},
        {"role": "user", "content": "Great city, but the rain takes adjustment"}
    ],
    user_id="user_123"
)
# Extracts: ["Lives in Seattle", "Likes Seattle", "Rain requires adjustment"]

# Semantic search
results = memory.search("Where does the user live?", user_id="user_123")

# Retrieve all
all_memories = memory.get_all(user_id="user_123")

# Cleanup
memory.close()

Async API

from remina import AsyncMemory
import asyncio

async def main():
    memory = AsyncMemory()
    await memory.add("I prefer dark mode", user_id="user_123")
    results = await memory.search("user preferences", user_id="user_123")
    await memory.close()

asyncio.run(main())

Configuration

from remina import Memory

config = {
    "llm": {
        "provider": "openai",  # or "anthropic", "gemini", "ollama"
        "config": {"model": "gpt-4o-mini"}
    },
    "embedder": {
        "provider": "openai",
        "config": {"model": "text-embedding-3-small"}
    },
    "vector_store": {
        "provider": "qdrant",  # or "pinecone", "chroma", "pgvector"
        "config": {"url": "http://localhost:6333"}
    },
    "storage": {
        "provider": "postgres",  # or "mongodb", "sqlite"
        "config": {
            "host": "localhost",
            "database": "remina",
            "user": "postgres",
            "password": "password"
        }
    },
    "cache": {
        "redis_url": "redis://localhost:6379",
        "enabled": True
    }
}

memory = Memory(config)

Supported Providers

Category	Providers
LLMs	OpenAI, Anthropic, Gemini, Ollama
Embeddings	OpenAI, Gemini, Cohere, Ollama, HuggingFace
Vector Stores	Pinecone, Qdrant, Chroma, pgvector
Storage	PostgreSQL, MongoDB, SQLite
Cache	Redis

Roadmap

v0.2 — Extraction & Intelligence

• Conversation ingestion pipeline
• Contradiction detection
• Memory consolidation

v0.3 — Graph & Relationships

• Memory graph with relationship linking
• Graph-enhanced retrieval
• Adaptive decay rates

v0.4 — Provider Expansion

• Weaviate, Milvus vector stores
• DynamoDB, Supabase storage
• Voyage, HuggingFace embeddings
• vLLM local inference

v1.0 — Production Hardening

• TypeScript SDK
• REST API wrapper
• Prometheus metrics
• Distributed tracing

Contributing

git clone https://github.com/bikidsx/remina
cd remina-memory
pip install -e ".[dev]"
pytest

See CONTRIBUTING.md for details.

License

Apache 2.0