Skip to content

Graphbase Memories MCP

Graph-backed persistent memory for AI coding agents, exposed as an MCP server.

Agents (Claude, Codex, Gemini, and others) call 20 structured tools to read and write scoped memory into a Neo4j graph database. Memory survives across sessions, accumulates decisions and patterns over time, and surfaces the most relevant context when you need it.

Why graph memory?

Most agent memory is flat — a list of notes, a vector store of embeddings, or session summaries that drift away. graphbase organizes memory as a property graph:

  • Scopes (global / project / focus) keep cross-project knowledge separate from initiative-specific context.
  • Artifact types (sessions, decisions, patterns, context snippets, entity facts) give structure to what agents remember.
  • Graph edges ([:SUPERSEDES], [:CONFLICTS_WITH], [:PRODUCED]) make the lineage and relationships between memories explicit and queryable.
graph TD
    A["AI Agent<br/>(Claude / Codex / Gemini)"]
    B["MCP Server — FastMCP<br/>22 async tools · stdio JSON-RPC 2.0"]
    C["Business Logic Layer<br/>engines/"]
    I[("Neo4j 5<br/>Graph Store")]

    A -->|"stdio"| B
    B --> C
    C -->|"Bolt :7687"| I

What agents can do

Action Tool
Load context before reasoning retrieve_context
Check scope before reading/writing get_scope_state
Fast BM25 keyword lookup for a specific topic memory_surface
Save a session summary save_session
Save session + decisions + patterns in one call store_session_with_learnings
Save an architectural decision (with dedup) save_decision
Save a repeatable workflow pattern save_pattern
Save a free-form context snippet save_context
Upsert a named entity and its relationships upsert_entity_with_deps
Obtain a global-scope write token request_global_write_approval
Route a task to the right reasoning mode route_analysis
Run memory hygiene (detect duplicates, stale items) run_hygiene
Check for pending or failed saves get_save_status
Preview nodes approaching the staleness threshold memory_freshness
Register a service into a workspace register_service
List services active in a workspace list_active_services
Search memory across services search_cross_service
Create a cross-service knowledge link link_cross_service
Propagate a breaking change across services propagate_impact
Get workspace health metrics graph_health
Find contradicting cross-service links detect_conflicts

Requirements

  • Python 3.11+
  • Neo4j 5 Community (or Enterprise) — local or remote
  • An MCP-compatible agent host (Claude Code, Cursor, Cline, etc.)