Labsco
MCP SERVER

MCP Memory Graph

by RetroRobAI

Give an agent memory that knows which of two conflicting notes to trust.

Agent Memory
Summary
Old instructions stop outranking new ones.

Ordinary memory servers return whatever is semantically closest, which means a stale note can beat the correction that replaced it. Here every store is checked for conflicts first, memories carry an authority tier, and superseded ones are marked with a typed edge — so retrieval returns what is current rather than what is nearest.

What it is

A memory server for Claude Code and other MCP agents that adds three things to plain vector recall: authority weighting by priority tier, conflict detection before every store, and typed relationship edges between memories.

What you get
  • Memories stored with a priority tier — high, medium or low — each carrying the authority weight retrieval ranks by
  • Retrieval that weighs authority rather than semantic closeness alone
  • Conflicts checked before a store, and resolved by authority
  • Superseded memories marked as such and linked, instead of left to compete
  • Typed edges added between memories, and the related ones walked
  • Memories updated and deleted
Requirements

A recent Python — pip install mcp-memory-graph — and a path for the SQLite database it writes, set in the client configuration.

Setup effort

One command — pip install mcp-memory-graph