CortexGit

Persistent memory for LLM agents. Event sourcing + semantic retrieval.

The Problem

LLM agents are stateless. They forget context between sessions. They can't coordinate without explicit message passing. They have no audit trail.

The Solution

CortexGit is an in-process memory system. Write events, retrieve context, persist facts. Works with any LLM, any agent framework.

Installation

pip install cortexgit

Quick Start

import asyncio
from cortexgit import CortexGit
from anthropic import Anthropic

# Initialize memory (creates local SQLite database)
memory = CortexGit()
client = Anthropic()

async def my_agent(user_query):
    session_id = "session-1"
    agent_id = "my-agent"
    
    # Retrieve relevant context from memory
    context = await memory.get_context(
        goal=user_query,
        budget_tokens=4000,
        session_id=session_id
    )
    
    # Call Claude with context
    response = await client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=1024,
        system=f"You are a helpful agent. Memory: {context}",
        messages=[{"role": "user", "content": user_query}]
    )
    
    # Remember what happened
    await memory.log_event(
        session_id=session_id,
        agent_id=agent_id,
        event_type="agent",
        payload={"query": user_query, "response": response.content[0].text}
    )
    
    return response.content[0].text

# Use it
print(asyncio.run(my_agent("What is 2+2?")))

Documentation

• API Reference
• Architecture
• Getting Started
• Multi-Provider LLM Setup
• Contributing
• Changelog

Configuration

By default, CortexGit uses SQLite (sqlite+aiosqlite:///cortexgit.db) which requires no external server or setup.

To use PostgreSQL in production:

1. Install PostgreSQL and create a database (e.g. cortexgit).
2. Set the DATABASE_URL environment variable:

   DATABASE_URL="postgresql+asyncpg://postgres:password@localhost:5432/cortexgit"

Features

• ✅ Append-only event log (source of truth)
• ✅ Persistent entity registry with conflict detection
• ✅ Automatic snapshot generation
• ✅ Semantic retrieval over compressed memory
• ✅ Works with any LLM (Claude, GPT, local models)
• ✅ Single import, no server needed

License

MIT