membrane-ai 3.1.8

Enterprise AI privacy middleware — reversible PII anonymization.

Membrane

Zero-latency, reversible PII anonymization middleware for LLMs.

---

[!IMPORTANT] Membrane Enterprise
The open-source SDK is built for developers and rapid prototyping. For production deployments in highly-regulated environments requiring SOC2 compliance, AWS KMS / HashiCorp Vault integrations, centralized RBAC, and immutable Audit Logging, explore our Enterprise tier.
Request Enterprise Demo

Installation

pip install membrane-ai

Quick Start

Membrane intercepts your prompt, deterministically masks the PII locally, and seamlessly restores it from the LLM's response. It is specifically built to handle LLM paraphrasing.

from membrane import TrustLayer

layer = TrustLayer(provider="openai", api_key="sk-...")

# Membrane strips the PII, queries the LLM, and rehydrates the result automatically.
response = layer.call("My SSN is 123-45-6789. Send the report to [e-mail].")

print(response["final_response"]) 

Features Breakdown

Capability	Open Source (Free)	Enterprise
Core Anonymization	Regex & Presidio NLP	Advanced Custom NER Models
Provider Support	OpenAI, Anthropic, Gemini, Ollama	Bring-Your-Own-Model (VPC)
Key Management	Environment Variables	AWS KMS, Azure KeyVault, HashiCorp Vault
Audit Logging	Local File-based	Datadog, Splunk, Immutable S3 Export
Access Control	N/A	Granular RBAC, SSO/SAML
Scaling & State	In-Memory (Single Node)	Distributed Redis Architecture

Why Membrane?

• Zero Trust Security: Sensitive data never leaves your infrastructure. Masking happens entirely locally before the outbound network request to external APIs (OpenAI/Anthropic) is even constructed.
• Near-Zero Latency: Built for high-throughput streaming environments and rigorous production traffic without adding computational overhead.
• Deterministic Context Preservation: We don't just crudely mask data; we maintain contextual entity trackers. Even when an LLM radically paraphrases a response, we still resolve the identity deterministically and mathematically.