WARNING

I've become disillusioned with automated context management tools like this, as it's nearly impossible to control. After a while, I always have to manually clean up the mess or correct inappropriate LLM notes. Instead, I created a tool that gives the LLM agent access to dynamically loaded context. However, the context itself is user-created: https://github.com/n-r-w/agent-standards-mcp

KnowledgeGraph MCP Server

A simple way to give LLMs persistent memory across conversations. This server lets Claude or vscode remember information about you, your projects, and your preferences using a knowledge graph.

Key Features:

• Multiple Storage Backends: PostgreSQL (recommended) or SQLite (local file)

• Project Separation: Keep different projects isolated (auto-detected using prompts)

• Better Search: Find information with fuzzy search and pagination

Complete Setup Guide

Follow these steps in order to get the knowledge graph working with Claude:

Step 1: Choose Your Installation Method

Option A: NPX (Easiest - No download needed)

Option B: Docker

Step 2: Choose Your Database

SQLite (Default - No setup needed):

• No database installation required

• Database file created automatically in [you home folder]/.knowledge-graph/

• Perfect for personal use and most scenarios

• This is the default backend

PostgreSQL (For advanced users):

• Install PostgreSQL on your system

• Create a database: CREATE DATABASE knowledgegraph;

• Better for production use with multiple concurrent users

Step 3: Configure client

Claude Desktop

Edit your Claude Desktop configuration file:

Find your config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

If you chose NPX + SQLite (default and easiest):

Note: SQLite will automatically create the database in [you home folder]/.knowledge-graph/knowledgegraph.db. To use a custom location, add: "KNOWLEDGEGRAPH_SQLITE_PATH": "/path/to/your/database.db"

If you chose Docker + SQLite (default):

Note: The volume mount ensures your data persists between Docker runs. For custom paths, add: -e KNOWLEDGEGRAPH_SQLITE_PATH=/app/.knowledge-graph/custom.db

If you chose PostgreSQL:

VS Code

If you also want to use this with VS Code, add this to your User Settings (JSON) or create .vscode/mcp.json:

Using NPX + SQLite (default):

Using Docker (default SQLite):

Using Docker + PostgreSQL:

First, ensure your PostgreSQL database is set up:

Then configure VS Code:

Alternative Docker + PostgreSQL (if

Important Notes:

• Replace yourpassword with your actual PostgreSQL password

• Ensure the knowledgegraph database exists before starting

• If you get connection errors, try the alternative configuration above

• For troubleshooting Docker + PostgreSQL issues, see the Common Issues section

Step 4: Choose Your LLM System Prompts

Customization:

• Modify entity types based on your domain

• Adjust search strategies for your data patterns

• Add domain-specific tags and relation types

LLM Compatibility:

• All LLMs behave differently. For some, general instructions are enough, while others need to describe everything in detail

• Use LLM to explain why it didn't use the knowledge graph. Ask Explain STEP-BY-STEP why you didn't use the knowledge graph? DO NOT DO ANYTHING ELSE to get a detailed report and identify issues with the instructions.

Available Prompts:

• Knowledge Graph

• Task Management

• Code Quality

• All-in-One

• Maintenance knowledge graph

Step 5: Restart Claude Desktop (or VS Code)

Close and reopen Claude Desktop. You should now see "Knowledge Graph" in your available tools.

Step 6: Test It Works

Quick Test Commands for LLMs:

1. "Remember that I prefer morning meetings" → Creates preference entity

2. "John Smith works at Google as a software engineer" → Creates person + company + relation

3. "Find all people who work at Google" → Tests search and relations

4. "Mark the morning meetings preference as urgent" → Tests tagging

Note: The service includes comprehensive input validation to prevent errors. If you encounter any issues, check the Troubleshooting Guide for common solutions.

How It Works - LLM Power Features

The knowledge graph enables powerful queries through four interconnected concepts:

1. Entities - Your Knowledge Nodes

Store people, projects, companies, technologies as searchable entities.

Real Example - Project Management:

LLM Benefit: Find "all available team leads" instantly with tag search.

2. Relations - Enable Discovery Queries

Connect entities to answer complex questions like "Who works on what?"

Real Example - Team Structure:

LLM Benefit: Query "Find all projects Sarah leads" or "Who leads Project Alpha?"

3. Observations - Atomic Facts

Store specific, searchable facts about entities.

Real Examples - Actionable Information:

• "Available for urgent tasks" → Find available people

• "Uses React 18.2" → Find projects with specific tech

• "Deadline: March 15, 2024" → Find upcoming deadlines

4. Tags - Instant Filtering

Enable immediate status and category searches.

Real Examples - Project Workflow:

• ["urgent", "in-progress", "frontend"] → Find urgent frontend tasks

• ["completed", "bug-fix"] → Track completed bug fixes

• ["available", "senior"] → Find available senior staff

Configuration Options

Environment Variables

The server supports several environment variables for customization:

Database Configuration

• KNOWLEDGEGRAPH_STORAGE_TYPE: Database type (sqlite or postgresql, default: sqlite)

• KNOWLEDGEGRAPH_CONNECTION_STRING: Database connection string

• KNOWLEDGEGRAPH_SQLITE_PATH: Custom SQLite database path (optional)

• KNOWLEDGEGRAPH_PROJECT: Project identifier for data isolation (default: knowledgegraph_default_project)

Search Configuration

• KNOWLEDGEGRAPH_SEARCH_MAX_RESULTS: Maximum number of results to return from database searches (default: 100, max: 1000)

• KNOWLEDGEGRAPH_SEARCH_BATCH_SIZE: Batch size for processing large query arrays (default: 10, max: 50)

• KNOWLEDGEGRAPH_SEARCH_MAX_CLIENT_ENTITIES: Maximum number of entities to load for client-side search (default: 10000, max: 100000)

• KNOWLEDGEGRAPH_SEARCH_CLIENT_CHUNK_SIZE: Chunk size for processing large datasets in client-side search (default: 1000, max: 10000)

Note: Search limits are automatically validated and clamped to safe ranges to prevent performance issues.

Performance Optimization

The search system includes several performance optimizations:

Entity Loading Limits:

• KNOWLEDGEGRAPH_SEARCH_MAX_CLIENT_ENTITIES limits how many entities are loaded for client-side search

• Prevents memory issues with large datasets

• Warning logged when limit is reached

• Applies to both SQLite and PostgreSQL backends

Chunked Processing:

• KNOWLEDGEGRAPH_SEARCH_CLIENT_CHUNK_SIZE controls chunk size for large entity sets

• Automatically used when entity count exceeds chunk size

• Improves memory usage and search performance

• Maintains result accuracy with deduplication

Recommended Values by Dataset Size:

• Small (< 1,000 entities): Default values work well

• Medium (1,000 - 10,000 entities): Consider KNOWLEDGEGRAPH_SEARCH_MAX_CLIENT_ENTITIES=5000, KNOWLEDGEGRAPH_SEARCH_CLIENT_CHUNK_SIZE=500

• Large (> 10,000 entities): Use database-level search when possible, or KNOWLEDGEGRAPH_SEARCH_MAX_CLIENT_ENTITIES=2000, KNOWLEDGEGRAPH_SEARCH_CLIENT_CHUNK_SIZE=200

Performance Monitoring:

• Warnings logged when limits are applied

• Chunking automatically logged for transparency

• Configuration validation prevents suboptimal settings

Available Tools

The server provides these tools for managing your knowledge graph:

Data Creation Tools

create_entities

CREATE new entities (people, concepts, objects) in knowledge graph.

• WHEN: Use for entities that don't exist yet

• CONSTRAINT: Each entity MUST have ≥1 non-empty observation

• BEHAVIOR: Ignores entities with existing names (use add_observations to update)

Input:

• entities (Entity[]): Array of entity objects. Each REQUIRES:

  • name (string): Unique identifier, non-empty

  • entityType (string): Category (e.g., 'person', 'project'), non-empty

  • observations (string[]): Facts about entity, MUST contain ≥1 non-empty string

  • tags (string[], optional): Exact-match labels for filtering

• project_id (string, optional): Project name to isolate data

create_relations

CONNECT entities to enable powerful queries and discovery.

• IMMEDIATE BENEFITS: Find all people at a company, all projects using a technology, all dependencies

• CRITICAL FOR: Team structures, project dependencies, technology stacks

• EXAMPLES: 'John works_at Google', 'React depends_on JavaScript', 'Project_Alpha managed_by Sarah'

Input:

• relations (Relation[]): Array of relationship objects. Each REQUIRES:

  • from (string): Source entity name (must exist)

  • to (string): Target entity name (must exist)

  • relationType (string): Relationship type in active voice (works_at, manages, depends_on, uses)

• project_id (string, optional): Project name to isolate data

add_observations

ADD factual observations to existing entities.

• REQUIREMENT: Target entity must exist, ≥1 non-empty observation per update

• BEST PRACTICE: Keep observations atomic and specific

Input:

• observations (ObservationUpdate[]): Array of observation updates. Each REQUIRES:

  • entityName (string): Target entity name (must exist)

  • observations (string[]): New facts to add, MUST contain ≥1 non-empty string

• project_id (string, optional): Project name to isolate data

add_tags

ADD status/category tags for INSTANT filtering.

• IMMEDIATE BENEFIT: Find entities by status (urgent, completed, in-progress) or type (technical, personal)

• REQUIRED: For efficient project management and quick retrieval

• EXAMPLES: ['urgent', 'completed', 'bug', 'feature', 'personal']

Input:

• updates (TagUpdate[]): Array of tag updates. Each REQUIRES:

  • entityName (string): Target entity name (must exist)

  • tags (string[]): Status/category tags to add (exact-match, case-sensitive)

• project_id (string, optional): Project name to isolate data

Data Retrieval Tools

read_graph

RETRIEVE complete knowledge graph with all entities and relationships.

• USE CASE: Full overview, understanding current state, seeing all connections

• SCOPE: Returns everything in specified project

Input:

• project_id (string, optional): Project name to isolate data

search_knowledge

SEARCH entities by text or tags. SUPPORTS MULTIPLE QUERIES for batch searching.

• MANDATORY STRATEGY: 1) Try searchMode='exact' first 2) If no results, use searchMode='fuzzy' 3) If still empty, lower fuzzyThreshold to 0.1

• EXACT MODE: Perfect substring matches (fast, precise)

• FUZZY MODE: Similar/misspelled terms (slower, broader)

• TAG SEARCH: Use exactTags for precise category filtering

• MULTIPLE QUERIES: Search for multiple objects in one call with automatic deduplication

Input:

• query (string | string[], optional): Search query for text search. Can be a single string or array of strings for multiple object search. OPTIONAL when exactTags is provided for tag-only searches.

• searchMode (string, optional): "exact" or "fuzzy" (default: "exact"). Use fuzzy only if exact returns no results

• fuzzyThreshold (number, optional): Fuzzy similarity threshold. 0.3=default, 0.1=very broad, 0.7=very strict. Lower values find more results

• exactTags (string[], optional): Tags for exact-match searching (case-sensitive). Use for category filtering

• tagMatchMode (string, optional): For exactTags: "any"=entities with ANY tag, "all"=entities with ALL tags (default: "any")

• page (number, optional): Page number for pagination (0-based, default: 0)

• pageSize (number, optional): Number of results per page (1-1000, default: 50)

• project_id (string, optional): Project name to isolate data

Examples:

• Basic search: search_knowledge(query="JavaScript", searchMode="exact")

• Paginated search: search_knowledge(query="React", page=0, pageSize=20)

• Large dataset: search_knowledge(query="components", page=2, pageSize=100)

• Multiple queries: search_knowledge(query=["JavaScript", "React"], page=0, pageSize=30)

• Tag + pagination: search_knowledge(query="React", exactTags=["frontend"], page=1, pageSize=25)

• Tag-only search: search_knowledge(exactTags=["urgent", "bug"], tagMatchMode="all") - NO QUERY NEEDED

Pagination Benefits:

• Performance: Database-level pagination with OFFSET/LIMIT for efficient large dataset handling

• Memory: Reduces memory usage by limiting results per request

• Navigation: Pagination metadata provides totalPages, currentPage, and navigation hints

• Scalability: Handles knowledge graphs with thousands of entities efficiently

open_nodes

RETRIEVE specific entities by exact names with their interconnections.

• RETURNS: Requested entities plus relationships between them

• USE CASE: When you know exact entity names and want detailed info

Input:

• names (string[]): Array of entity names to retrieve

• project_id (string, optional): Project name to isolate data

Data Management Tools

delete_entities

PERMANENTLY DELETE entities and all their relationships.

• WARNING: Cannot be undone, cascades to remove all connections

• USE CASE: Entities no longer relevant or created in error

Input:

• entityNames (string[]): Array of entity names to delete

• project_id (string, optional): Project name to isolate data

delete_observations

REMOVE specific observations from entities while keeping entities intact.

• USE CASE: Correct misinformation or remove obsolete details

• PRESERVATION: Entity and other observations remain unchanged

Input:

• deletions (ObservationDeletion[]): Array of deletion requests. Each REQUIRES:

  • entityName (string): Target entity name

  • observations (string[]): Specific observations to remove

• project_id (string, optional): Project name to isolate data

delete_relations

UPDATE relationship structure when connections change.

• CRITICAL FOR: Job changes (remove old 'works_at'), project completion (remove 'assigned_to'), technology migration (remove old 'uses')

• MAINTAINS: Accurate network structure and prevents confusion

• WORKFLOW: Always remove outdated relations when creating new ones

Input:

• relations (Relation[]): Array of relations to delete. Each REQUIRES:

  • from (string): Source entity name

  • to (string): Target entity name

  • relationType (string): Exact relationship type to remove

• project_id (string, optional): Project name to isolate data

remove_tags

UPDATE entity status by removing outdated tags.

• CRITICAL: For status tracking - remove 'in-progress' when completed, 'urgent' when resolved

• MAINTAINS: Clean search results and accurate status

• WORKFLOW: Always remove old status tags when adding new ones

Input:

• updates (TagUpdate[]): Array of tag removal requests. Each REQUIRES:

  • entityName (string): Target entity name

  • tags (string[]): Outdated tags to remove (exact-match, case-sensitive)

• project_id (string, optional): Project name to isolate data

Development and Testing

Multi-Backend Testing

This project includes comprehensive multi-backend testing to ensure compatibility across both SQLite and PostgreSQL:

Run tests against both backends:

Run all tests (original + multi-backend):

Using Taskfile (if installed):

Development Setup

Clone and setup:

Run tests:

Troubleshooting

If you encounter any issues during setup or usage, please refer to our comprehensive Troubleshooting Guide, which covers:

• Input validation errors

• Database connection problems

• Configuration issues

• Docker-related challenges

• Test execution failures

• Performance optimization

The guide includes step-by-step solutions for common problems and diagnostic commands to help identify issues.

Based on MCP Memory Server

This is an enhanced version of the official MCP Memory Server with additional features:

• Multiple Storage Options: PostgreSQL (recommended) or SQLite (local file)

• Project Separation: Keep different projects isolated

• Better Search: Find information with fuzzy search

• Easy Setup: Docker support and simple installation

License

MIT License - Feel free to use, modify, and distribute this software.