LangGraph Memory Filter for Open WebUI

A sophisticated persistent memory system for Open WebUI that extracts and maintains user information across conversations using LangGraph and PostgreSQL.

Author

Beau D'Amore
www.damore.ai

✨ Features

• Persistent Memory: Remembers user preferences, relationships, goals, interests, and personal information
• PII Protection: 3-layer defense-in-depth filtering prevents storage of phone numbers, SSNs, addresses, credit cards, and other personally identifiable information — while still storing personality traits, preferences, and ownership types
• Preference Evolution Tracking: Tracks how user preferences change over time (e.g., "used to like coffee → now prefers tea")
• PostgreSQL Backend: Reliable, persistent storage with LangGraph checkpoint support
• Multi-Model Support: Works with GPT-4, Llama, Qwen, Gemini, and other extraction models
• Schema Versioning: Built-in migration system for data structure evolution

📁 Project Structure

memory_langgraph/
├── README.md                    # This file
├── DEPENDENCY_RESOLUTION.md     # Dependency troubleshooting guide
├── test_imports.py              # Dependency verification script
├── filter/
│   ├── langgraph_memory_filter.py    # Main filter code
│   └── requirements-langgraph.txt    # Python dependencies
├── docker/
│   └── docker-compose.langgraph.yml  # PostgreSQL container
├── docs/
│   ├── LANGGRAPH_SETUP.md            # Full installation guide
│   ├── EXTRACTION_MODEL_SETUP.md     # Extraction model configuration
│   ├── AI_DEVELOPMENT_GUIDE.md       # Guide for AI assistants
│   └── ...
└── prompt/
    └── EXTRACTION_PROMPT.md          # Extraction model system prompt

🚀 Quick Start

1. Start PostgreSQL

cd docker
docker-compose -f docker-compose.langgraph.yml up -d

2. Install Dependencies

# In your Open WebUI container
docker exec -it <open-webui-container> pip install \
  "langgraph>=1.0.0" \
  langgraph-checkpoint-postgres \
  "psycopg[binary]" \
  psycopg-pool

3. Install the Filter

1. Go to Open WebUI Admin Panel → Functions
2. Click "+ Add Function"
3. Paste the contents of filter/langgraph_memory_filter.py
4. Save and enable the filter

4. Configure Extraction Model

See docs/EXTRACTION_MODEL_SETUP.md for detailed instructions.

Quick version:

1. Create a custom model in Open WebUI with the system prompt from prompt/EXTRACTION_PROMPT.md
2. Set the filter's extraction_model_id valve to your model's ID

📖 Documentation

Document	Description
LANGGRAPH_SETUP.md	Complete installation and setup guide
EXTRACTION_MODEL_SETUP.md	Extraction model configuration
AI_DEVELOPMENT_GUIDE.md	Guide for AI assistants modifying this code
DEPENDENCY_RESOLUTION.md	Dependency troubleshooting

🔧 Requirements

• Open WebUI: Latest version
• Python: 3.10+
• PostgreSQL: 16+ (provided via Docker)
• Dependencies:

  langgraph>=1.0.0
  langgraph-checkpoint-postgres
  psycopg[binary]
  psycopg-pool
  

🐳 Docker Setup

PostgreSQL Container

cd docker
docker-compose -f docker-compose.langgraph.yml up -d

Default Connection:

• Host: localhost (or langgraph-postgres from other containers)
• Port: 5432
• Database: langgraph_memory
• User: langgraph
• Password: langgraph_password_change_me

⚙️ Configuration

The filter has configurable "valves" (settings):

Valve	Default	Description
enable_filter	true	Enable/disable the filter
postgres_connection_string	postgresql://...	Database connection
extraction_model_id	""	Model ID for memory extraction
debug_logging	false	Enable verbose logging
enable_memory_injection	true	Inject memories into context
pii_filter_enabled	true	Enable PII detection and filtering
pii_filter_mode	"remove"	remove drops facts with PII; redact stores with [REDACTED]
pii_scrub_input	true	Scrub PII from messages before sending to extraction model
pii_patterns_enabled	all 10	Which PII types to detect (see PII Protection section)

See PII Protection for details on privacy filtering.

🧪 Testing

Verify Dependencies

python test_imports.py

Test Database Connection

docker exec -it langgraph-postgres psql -U langgraph -d langgraph_memory -c "SELECT 1;"

View Stored Memories

docker exec -it langgraph-postgres psql -U langgraph -d langgraph_memory -c \
  "SELECT thread_id, checkpoint FROM checkpoints ORDER BY thread_id LIMIT 5;"

� PII Protection

The filter includes a 3-layer defense-in-depth system to prevent storage of personally identifiable information:

Layer 1: Prompt Guardrails

The extraction model's system prompt explicitly instructs it to never extract PII. Examples are provided showing how to extract non-PII meaning from PII-containing messages (e.g., "I bought a house at 123 Oak Lane" → store ownership: house, recently purchased a house).

Layer 2: Regex Pre-Scrub

Before messages reach the extraction model, all PII is replaced with [REDACTED]. The LLM never sees raw PII data.

Layer 3: Post-Extraction Validation

Every extracted fact is scanned for PII in both subject and value fields, plus a subject blocklist. Facts with PII are either dropped or redacted based on config.

Detected PII Types

Pattern	Example
Social Security Number	123-45-6789
Credit Card	4111-1111-1111-1111
Phone Number	(555) 123-4567
Email Address	user@example.com
Street Address	123 Main Street Apt 4B
Passport Number	Passport: A12345678
Driver's License	Driver's license: D1234567
Bank Account	Account #: 123456789
Date of Birth	DOB: 01/15/1990
IP Address	192.168.1.1

What IS Stored

• Personality traits, preferences, opinions
• Ownership types ("owns a Tesla" — not the VIN)
• Relationships ("married to Sarah" — not her phone number)
• Goals, skills, interests, hobbies
• Professional info ("engineer at Acme" — not employee ID)

📊 Schema Version

Current: Schema v4 (LLM-Powered Semantic Merge + PII Protection)

Schema v4 features:

• LLM-powered semantic merge replaces code-based deduplication
• Flexible fact-based storage (replaces rigid type system)
• 3-layer PII filtering (prompt guardrails, regex pre-scrub, post-extraction validation)
• schema_migrations table for version tracking

🤖 For AI Assistants

If you're an AI assistant modifying this codebase, please read:

📚 docs/AI_DEVELOPMENT_GUIDE.md

This document contains critical information about:

• Schema version requirements
• Pitfalls to avoid
• Code modification procedures
• Testing requirements

🐛 Troubleshooting

Common Errors

cannot import name 'RunnableSerializable'

pip install "langgraph>=1.0.0"

ImportError: no pq wrapper available

pip install "psycopg[binary]"

ModuleNotFoundError: No module named 'psycopg_pool'

pip install psycopg-pool

See DEPENDENCY_RESOLUTION.md for detailed troubleshooting.

📝 License

MIT License

🙏 Acknowledgments

• Open WebUI - The amazing AI chat interface
• LangGraph - State machine for LLM applications
• LangChain - LLM application framework