j5ed-knowledge-graph

A knowledge graph memory server for the Model Context Protocol (MCP). Gives AI assistants persistent, structured memory across sessions using a simple entity-relation-observation graph.

Forked from the official @modelcontextprotocol/server-memory and significantly enhanced with multi-agent isolation, tiered search, index navigation, backup rotation, and three transport modes.

What's different from upstream

Related MCP server: Knowledge Master

Installation

npm install j5ed-knowledge-graph

Or run directly:

npx j5ed-knowledge-graph

MCP Configuration

stdio (local, default)

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "j5ed-knowledge-graph"],
      "env": {
        "MEMORY_FILE_PATH": "/path/to/memory.json"
      }
    }
  }
}

Streamable HTTP (network)

npx j5ed-knowledge-graph --http --port 3100

{
  "mcpServers": {
    "memory": {
      "url": "http://localhost:3100/mcp"
    }
  }
}

SSE (legacy network)

npx j5ed-knowledge-graph --sse --port 3100

Environment Variables

Multi-Agent Setup

When multiple agents share a graph, each agent's writes are tagged with authored_by:<agentId> and reads are filtered to show only entities owned by (or granted to) that agent.

Example: two agents with cross-read access

AGENT_CREDENTIALS="assistant:token-abc,researcher:token-xyz" \
AGENT_READ_GRANTS="assistant:researcher" \
npx j5ed-knowledge-graph --http --port 3100

• assistant authenticates with Bearer token-abc, sees its own entities + researcher's

• researcher authenticates with Bearer token-xyz, sees only its own entities

• Writes are always scoped — agents can only delete what they authored

Tools

Index Navigation Pattern

The graph supports an index-first navigation pattern for large knowledge bases:

1. read_graph() — returns lightweight stubs for index entities (name, type, summary)

2. Pick relevant index — open_nodes(["MY_INDEX"])

3. See inbound relations — entities with indexed_in relations to the index are its table of contents

4. Drill into specifics — open_nodes(["SPECIFIC_ENTITY"])

Index entities are detected by entityType containing "index" (case-insensitive) or entity name ending with _INDEX.

Search Behavior

Multi-word queries are tokenized. Results are grouped into tiers by how many tokens matched:

• Tier 1: entities matching any 1 token (broadest)

• Tier N: entities matching all N tokens (most specific)

Within each tier, entities are ranked by field priority: index entities first, then name matches (+30), type matches (+20), observation matches (+10). Per-tier caps prevent noise.

Results are returned as lightweight stubs (name, type, matchedIn fields, optional snippet). Use open_nodes() to get full observations.

Storage Format

NDJSON (newline-delimited JSON). Each line is either:

{"type":"entity","name":"MY_ENTITY","entityType":"service","observations":["fact 1","fact 2"]}
{"type":"relation","from":"MY_ENTITY","to":"MY_INDEX","relationType":"indexed_in"}

Docker

docker build -t j5ed-knowledge-graph .
docker run -i -v graph-data:/app/dist --rm j5ed-knowledge-graph

Development

git clone https://github.com/arktechnwa/j5ed-knowledge-graph.git
cd j5ed-knowledge-graph
npm install
npm run build
npm test        # 7 suites, 80 tests
npm run dev     # stdio mode with ts-node

License

MIT