🧠 MemFS

A knowledge graph management system based on MCP server-memory, deeply refactored with filesystem-inspired design

💡 Acknowledgments: Original @modelcontextprotocol/server-memory
Inspired by it, though heavily reimagined.

🎯 One-Line Description

Bringing modern filesystem concepts to knowledge graph management, combined with BM25 + fuzzy search for intelligent retrieval, designed for LLM-assisted humanities and social sciences research.

📖 中文版文档: docs/README_zh-CN.md

🚀 Quick Start

Prerequisites

# Check Node.js version
node --version  # Must be v22.0.0 or higher

Installation & Run

Quickest way (npx):

npx @qty/memfs

Or clone and run:

# 1. Clone or download the project
cd MemFS

# 2. Install dependencies
npm install

# 3. Run server
node index.js

# Or specify custom storage path
MEMORY_DIR=~/my-knowledge
MEMORY_FILE_PATH=./custom/memory.jsonl

Configure as MCP Server

OpenCode format:

{
  "mcpServers": {
    "memory": {
      "type": "local",
      "command": ["npx", "-y", "@qty/memfs"],
      "enabled": true
    }
  }
}

VSCode / ClaudeCode / Cherry Studio / AstrBot format:

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@qty/memfs"],
      "enabled": true
    }
  }
}

📖 Core Concepts

| Concept | Description | Analogy | |---------|-------------|---------| | Entity | Nodes in the knowledge graph | File | | Observation | Properties/descriptions of entities | inode | | Relation | Connections between entities | Soft link | | Reference | Pointers from entities to observations | Hard link |

💡 Core Design Philosophy

1. Transformer-Ready: On-Demand Retrieval

flowchart TD
    LLM["LLM"]
    ATT["Attention Mechanism"]
    MCP["MCP Protocol"]
    MEM["MemFS\nOn-demand structured data"]

    LLM --> ATT --> MCP --> MEM
    MEM -.-> |Returns results| LLM

Core principle: Don't stuff all knowledge into context—retrieve on demand.

2. Lightweight Design

| Dimension | Traditional Solution | MemFS | |-----------|---------------------|-------| | Deployment | Database + Vector Engine | Pure Node.js | | Resources | GPU recommended, high memory | CPU only | | Explainability | Black-box models | BM25 transparent & controllable |

3. Local JSONL Storage

{"type":"entity","name":"Weber","entityType":"person","definition":"German sociologist","observationIds":[1,2]}
{"type":"observation","id":1,"content":"Author of 'The Protestant Ethic'","createdAt":{"utc":"2026-02-08T13:53:07Z","timezone":"Asia/Shanghai"}}
{"type":"relation","from":"Weber","to":"Durkheim","relationType":"contemporary"}

Advantages: Editable with any text editor, Git-version-controllable, printable.

4. Humanities & Social Sciences Customization

| Requirement Type | Traditional | MemFS | |-----------------|-------------|-------| | Knowledge units | Functions/Classes | Concepts/People/Documents | | Relationship types | Function calls | Influence/Reference/Comparison | | Update frequency | High-frequency | Low-frequency add, high-frequency reference |

📦 Complete API Tools (16 total)

Create

| Tool | Function | Example | |------|----------|---------| | createEntity | Batch create entities (with observations) | Add concepts, people, documents | | createRelation | Create relations between entities | Mark references, comparisons, influences | | addObservation | Add observations to existing entities | Supplement reading notes |

Read

| Tool | Function | Example | |------|----------|---------| | searchNode | BM25 + Fuzzy hybrid search | Intelligent knowledge search | | readNode | Read complete entity information | Get detailed attributes and relations | | readObservation | Batch read observations by ID | Verify specific observations | | listNode | List all entity overviews | Browse knowledge structure | | listGraph | Read entire knowledge graph | Batch export, migration | | howWork | Get recommended workflow guidance | Learn how to use the system |

Update

| Tool | Function | Example | |------|----------|---------| | updateNode | Update entities and observations (Copy-on-Write) | Modify definitions, update notes | | updateObservation | Batch update observation content | Batch correct information |

Delete

| Tool | Function | Example | |------|----------|---------| | deleteEntity | Delete entities and relations | Remove outdated entries | | deleteRelation | Delete specific relations | Unlink entities | | deleteObservation | Unlink observations (preserve observation) | Remove references | | getOrphanObservation | Find orphan observations | Discover invalid data | | recycleObservation | Permanently delete observations | Clean up unused data |

🔍 Hybrid Search (searchNode)

Core Features

| Feature | Description | |---------|-------------| | BM25 | Considers term frequency and document frequency | | Fuzzy Search | Tolerates typos, supports approximate matching | | Query Tokenization | Tokenize → Search individually → Aggregate → Deduplicate | | Weighted Fusion | BM25 0.7 + Fuzzy 0.3, combined ranking |

Parameters

// Default hybrid search
await searchNode("functionalism");  // BM25 + Fuzzy

// Traditional keyword search
await searchNode("functionalism", { basicFetch: true });

// Custom parameters
await searchNode("sociology", {
    limit: 15,          // Return count
    bm25Weight: 0.7,    // BM25 weight
    fuzzyWeight: 0.3,   // Fuzzy search weight
    minScore: 0.01      // Minimum relevance threshold
});

Field Weights

| Field | Weight | Description | |-------|--------|-------------| | name | 5.0 | Highest - entity name | | entityType | 4.0 | Entity type | | definition | 4.0 | Definition description | | observation | 3.0 | Observation content |

🔧 Filesystem-Inspired Design

Architecture Analogy

| Filesystem Concept | MemFS Implementation | Solves | |-------------------|---------------------|--------| | Inode Table | Centralized observation storage | Data redundancy | | Hard Links | Multiple entities reference same observation | Shared reuse | | Soft Links | Entity relations | Flexible associations | | Copy-on-Write | Copy-on-Write updates | Concurrency safety | | Orphan Detection | Orphan observation cleanup | Resource recovery |

Observation Sharing

// Create two entities sharing the same observation
await createEntity([
  { name: "Zhang San", observations: ["Programmer"] },
  { name: "Li Si", observations: ["Programmer"] }
]);

// Under the hood: same observation ID is reused
{
  entities: [
    { name: "Zhang San", observationIds: [1] },
    { name: "Li Si", observationIds: [1] }
  ],
  observations: [
    { id: 1, content: "Programmer" }
  ]
}

Copy-on-Write

// Update a shared observation
await updateNode({
  entityName: "Zhang San",
  observationUpdates: [
    { oldContent: "Programmer", newContent: "Senior Programmer" }
  ]
});

// Result: Zhang San gets new observation, Li Si keeps original
{
  observations: [
    { id: 1, content: "Programmer" },      // Li Si uses
    { id: 2, content: "Senior Programmer" } // Zhang San's new observation
  ]
}

📁 Data Format

JSONL Storage

{"type":"entity","name":"Weber","entityType":"person","definition":"German sociologist","definitionSource":"Wikipedia","observationIds":[1,2]}
{"type":"observation","id":1,"content":"Author of 'The Protestant Ethic'","createdAt":{"utc":"2026-02-08T13:53:07Z","timezone":"Asia/Shanghai"}}
{"type":"observation","id":2,"content":"Contemporary with Durkheim and Marx","createdAt":{"utc":"2026-02-08T14:00:00Z","timezone":"Asia/Shanghai"}}
{"type":"relation","from":"Weber","to":"Durkheim","relationType":"contemporary"}

Storage Locations

| Method | Path | |--------|------| | Default | ~/.memory/memory.jsonl | | Custom directory | MEMORY_DIR=/path/to/data | | Custom path | MEMORY_FILE_PATH=/path/to/file.jsonl |

🧪 Testing

# Full test suite (45 tests)
node test_full.mjs

# Hybrid search tests (38 tests)
node test_hybrid_search.mjs

# Observation search tests
node test_observation_search.mjs

⚙️ Comparison with Original MCP Memory

| Dimension | Original | MemFS | |-----------|----------|-------| | Observation Storage | Embedded in entities | Centralized + ID reference | | Data Sharing | Not supported | Hard-link style sharing | | Update Mechanism | Direct overwrite | Copy-on-Write | | Search Capability | Simple keyword | BM25 + Fuzzy | | Orphan Detection |理论上不存在孤儿观察 | Supported | | Cache Mechanism | None | 30s TTL | | Windows Compatibility | Unknown | Graceful degradation |

📚 Design Philosophy

What? You're still reading? Well, alright.

Honestly, this project started because:

1. LLM context is limited — can't stuff all knowledge into prompts
2. Filesystem is a great invention — handling "multiple data sharing same content" is mature
3. Humanities research has special needs — concepts, literature, citation relationships
4. Controllability > SOTA — no need for black-box vector models

So:

• Borrow filesystem wisdom: inode table, hard links, copy-on-write
• Search uses BM25 + Fuzzy: lightweight, explainable, transparent, controllable
• Expose as tools: 16 MCP tools, LLM calls on demand

Result? — A quiet, efficient, unobtrusive knowledge management tool.

📄 License

Apache License 2.0

Manage knowledge the filesystem way—bringing order to chaos.