om-memory 0.2.2

Human-like memory for AI agents. 10x cheaper than RAG. Zero vector DB needed.

om-memory

Human-like memory for AI agents. 10x cheaper than RAG. Zero vector DB needed.

om-memory is the first-ever Python implementation of Observational Memory (OM) — a revolutionary approach to AI agent memory. Instead of embedding every message and retrieving similar ones per turn (RAG), OM continuously compresses conversation history into a dense, evolving text log of "observations" using two background agents (Observer and Reflector).

What is Observational Memory?

Observational Memory (OM) maintains a stable text context window of the user's conversation. The observation log is plain text that stays stable in the context window, enabling prompt caching (75-90% token cost discount from providers like OpenAI/Anthropic). Traditional RAG injects different retrieved chunks every turn, breaking the cache and costing up to 10x more.

Why om-memory?

• 10x Cheaper than RAG: By leveraging prompt caching on a stable context window.
• Zero Vector DB Needed: Uses standard storage backends (SQLite, Postgres, etc.) — no embeddings, no vector search.
• Better Accuracy: Maintains narrative continuity better than fragmented vector retrieval (Highest recorded on LongMemEval).
• Framework-Agnostic: Middleware pattern. om-memory manages context, you make your own LLM calls with LangChain, LlamaIndex, or raw Python.
• Observable by Default: Full event tracking, metrics, and Streamlit dashboard integration.

Quick Start

om-memory provides sensible defaults out of the box (SQLite storage, OpenAI gpt-4o-mini for background compression).

import asyncio
from om_memory import ObservationalMemory

async def main():
    # 1. Initialize OM (Zero config needed, uses SQLite + OPENAI_API_KEY)
    om = ObservationalMemory()
    
    # 2. Get context for a user thread
    thread_id = "user_123"
    context = await om.aget_context(thread_id=thread_id)
    
    # 3. Build your prompt and call YOUR LLM
    prompt = f"System: You are a helpful assistant.\n{context}\nUser: Hello!"
    response = "Hello! How can I help you today?" # Replace with actual LLM call
    
    # 4. Tell OM what happened so it can remember for next time
    await om.aadd_message(thread_id=thread_id, role="user", content="Hello!")
    await om.aadd_message(thread_id=thread_id, role="assistant", content=response)

if __name__ == "__main__":
    asyncio.run(main())

How It Works

1. Block 1 (Observations): A compressed, timestamped log of facts, decisions, and preferences. Handled by the Reflector agent.
2. Block 2 (Recent Messages): The uncompressed recent turns of the conversation. Once this grows past a threshold, the Observer agent compresses it into Block 1.

Installation

pip install om-memory

Optional dependencies:

• pip install om-memory[anthropic] - Anthropic provider support
• pip install om-memory[postgres] - PostgreSQL storage backend
• pip install om-memory[dashboard] - Streamlit dashboard

Documentation

Full documentation available in the repository examples.

License

Apache 2.0