Graphmind Context Graphs — Python SDK

A "Director's Commentary" middleware for LangChain Python agents. Captures why agents make decisions, not just what they did.

Installation

pip install graphmind-context-graphs

Prerequisites

# Graphmind database
docker run -d --name graphmind -p 8080:8080 fabischk/graphmind:latest

# LangChain + provider
pip install langchain langchain-core langgraph langchain-openai

Quick Start

from langchain.agents import create_agent
from graphmind_context_graphs import (
    create_context_graph,
    ContextGraphConfig,
    EmbeddingConfig,
)

# Your embedding provider (must implement embed, embed_batch, dimensions)
embedding_provider = MyEmbeddingProvider()

# Initialize
cg = create_context_graph(ContextGraphConfig(
    tenant="my_company",
    project="support",
    agent="support-agent",
    embedding=EmbeddingConfig(provider=embedding_provider, dimensions=1536),
    observer_model="openai:gpt-4.1-mini",  # Optional: for LLM-powered extraction
))

# Create agent with middleware + brain-mapping tools
agent = create_agent(
    "openai:gpt-4.1",
    tools=[*my_tools, *cg.tools],
    middleware=cg.middleware,
)

# Use — context captured and injected automatically
result = agent.invoke({"messages": [{"role": "user", "content": "My account is locked"}]})

# Evolve knowledge
cg.lifecycle.validate_trace(trace_id, ValidationResult(trace_id=trace_id, success=True))
cg.lifecycle.synthesize_rules()

Features

• Decision Trace Capture — Intent/Constraint/Action/Justification triplets
• Dynamic Brain Mapping — Agents create entities (CodeFile, Contract, etc.) and relationships
• Schema Awareness — Agent-scoped; only domain entities shown
• LLM-Powered Extraction — Observer model extracts domain, concepts, constraints
• Knowledge Lifecycle — Capture → Validate → Synthesize → Prune
• Multi-Agent — Shared, isolated, or selective context sharing
• Vector Search — Semantic similarity via Graphmind SEARCH clause

API Mirror

This SDK mirrors the TypeScript SDK. Both use the same Graphmind database and Cypher queries. Agents running in different languages can share context graphs.

License

Apache License 2.0