MCP Server

The MCP (Model Context Protocol) server is how pgmemory connects to any AI coding tool — Cursor, Windsurf, Cline, Claude Code, custom pipelines, or anything that speaks MCP. It's the universal integration point.

Setup

Add pgmemory to your tool's MCP configuration:

{
  "mcpServers": {
    "pgmemory": {
      "command": "pgmemory",
      "args": ["mcp"]
    }
  }
}

This works with Claude Code, Cursor, Windsurf, Cline, and any tool that supports MCP servers. The pgmemory daemon must be running (pgmemory start) for MCP tools to work.

Available tools

The MCP server provides 10 tools:

Core knowledge operations

Tool	What it does
memory_search	Search the knowledge base with a natural language query
memory_store	Store new knowledge (auto-deduplicated, noise-filtered, secrets scrubbed)
memory_list	Browse stored knowledge, optionally filtered by text
memory_delete	Remove a specific knowledge item

Source ingestion

Tool	What it does
source_ingest	Crawl a URL (wiki, docs site) and add its content to the knowledge base
source_upload	Upload files directly into memory
source_list	List all ingested sources and their status
source_remove	Remove a source and all its associated knowledge

Monitoring

Tool	What it does
quality_stats	Check the knowledge base health — retrieval counts, learning status
database_list	List connected databases and roles

How it fits together

Any AI tool ←→ MCP (stdin/stdout) ←→ pgmemory ←→ PostgreSQL + pgvector

The MCP server is the full knowledge management interface — read, write, update, and delete:

• Search (memory_search) — runs the retrieval pipeline with hybrid search, quality filtering, and diversity optimization
• Store (memory_store) — adds new knowledge through the capture pipeline with noise filtering, secret scrubbing, and deduplication
• Delete (memory_delete) — removes outdated or incorrect knowledge items
• Browse (memory_list) — lists stored knowledge

The proxy handles automatic capture from conversations. MCP handles everything else.

The good citizen pattern

AI tools that connect via MCP should actively maintain knowledge quality. When a tool retrieves knowledge via memory_search and discovers that the information is outdated, inaccurate, or no longer relevant, it should correct the record:

1. Delete the outdated item with memory_delete
2. Store the corrected version with memory_store

This is the expected behavior. AI tools using pgmemory should treat knowledge maintenance as part of their workflow.

Read-only usage

Any tool can connect via MCP with --read-only and only use memory_search — consuming knowledge without contributing or modifying:

{
  "mcpServers": {
    "pgmemory": {
      "command": "pgmemory",
      "args": ["mcp", "--read-only"]
    }
  }
}

See Read-Only Mode for more on this pattern.

For teams using multiple tools

A common team setup:

Team member	Tool	Connection	Contribution
Alice	Claude Code	Proxy + MCP	Automatic capture via proxy, retrieval + maintenance via MCP
Bob	Cursor	MCP	Agent searches, stores, and maintains knowledge
Carol	Custom pipeline	MCP (read-only)	Reads team knowledge, doesn't write or modify
Dave	Claude Code + Cursor	Proxy + MCP	Full capture + full knowledge management

All four draw from and contribute to the same PostgreSQL knowledge store.