Provides tools for storing and querying episodic memories, relationships, and graph structures using Neo4j, including Cypher read/write operations and bond evolution tracking.
Uses OpenAI embeddings for semantic vector search to retrieve memories based on similarity, enabling pattern completion and context-aware recall.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Hippocampal Memory MCP Encode a memory about our project kickoff meeting"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
An MCP server for neuroanatomically-inspired memory consolidation using Neo4j and semantic search.
Episodic Memory Encoding: Create memory events with emotional valence and temporal context
Semantic Retrieval: Vector similarity search using OpenAI embeddings
Graph Operations: Full Cypher query support for reading and writing
Relationship Tracking: Monitor bond strength evolution over time
Extensible Schema: Additional tools in hippocampus-extension.mjs ready for integration
Related MCP server: RAG Memory MCP
See CHECKLIST.md for daily startup instructions.
npm installYou need Neo4j Desktop running locally with a database on port 7687.
npm run setup-schemaStart ngrok to expose your local Neo4j:
ngrok tcp 7687Copy the tunnel URL (e.g., tcp://2.tcp.us-cal-1.ngrok.io:12841)
Edit %APPDATA%\Claude\claude_desktop_config.json:
{
"mcpServers": {
"hippocampal-memory": {
"command": "node",
"args": ["C:\\Users\\harve\\Neo4j\\hippocampal-mcp-server.mjs"],
"env": {
"NEO4J_URI": "bolt://YOUR_NGROK_URL",
"NEO4J_USER": "neo4j",
"NEO4J_PASSWORD": "your_password",
"OPENAI_API_KEY": "your_openai_key"
}
}
}
}The MCP server will connect automatically.
hippocampus_write_event - Structured event creation with Who/Why/What/Where/Effects blocks
hippocampus_write_reflection - Agent-relative memory slices with egocentric perspective
hippocampus_search_events - Pattern completion retrieval with temporal/participant/effect filters
encode_memory - Save episodic memories with vector embeddings
recall_memory - Semantic + temporal + emotional search
query_graph - Read-only Cypher queries
mutate_graph - Write operations (CREATE, MERGE, etc.)
evolve_bond - Track relationship dynamics over time
Current Setup:
Neo4j Desktop running locally on port 7687
ngrok tunnel for remote access
Claude Desktop connects via stdio transport
Render deployment at https://hippocampal-memory-mcp.onrender.com (HTTP/SSE transport)
Files:
hippocampal-mcp-server.mjs - Main server (stdio for Claude Desktop)
remote-mcp-server.mjs - HTTP/SSE server (for remote connections)
hippocampus-extension.mjs - Additional biomimetic tools
setup-vector-index.mjs - Schema initialization
test-connection.mjs - Local connection test
test-tunnel.mjs - Tunnel connection test
Core Nodes:
Event - Episodic memories with vector embeddings
Person - Human and AI entities
Project - Ongoing work
Concept - Abstract ideas
Place, Catalyst, Entity, Target, Effect, Reflection, Agent (extension schema)
Key Relationships:
INVOLVES - Event → Entity (with role & salience)
PRECEDED - Event → Event (causal chains)
CONSOLIDATED_TO - Event → Concept/Person/Project
BOND - Person ↔ Person (with strength trajectory)
PARTICIPATED_IN, CATALYZED_BY, HELD_AT, HAD_EFFECT_ON, etc. (extension schema)
Indexes:
Vector index on Event.embedding (1536 dimensions, cosine similarity)
Unique constraints on id fields for Person, Project, Concept, Event
ngrok URL changes on every restart (unless you pay for static URL)
Update NEO4J_URI in Claude Desktop config when ngrok URL changes
Neo4j Desktop must be running before starting Claude Desktop
Environment variables are in Claude Desktop config (NOT .env file)
Render deployment requires persistent Neo4j (Aura) - local tunnel won't work
Test local connection:
node test-connection.mjsTest tunnel connection:
node test-tunnel.mjsCHECKLIST.md - Daily startup and tool addition guide
neo4j-agent.instructions.md - Comprehensive Neo4j/Cypher/GraphQL guide
docs/ - Archived setup guides and deployment notes
Currently deployed to Render at: https://hippocampal-memory-mcp.onrender.com
For deployment details, see archived docs.
Create episodic memory event with emotional valence, temporal context, and entity involvement. Automatically generates embeddings for semantic retrieval.
Inputs: event (required), involves, precedes, consolidates_to.
Update bond strength between entities and track relationship trajectory over time. Automatically records milestones and calculates momentum.
Inputs: new_strength (required), to_entity_id (required), from_entity_id (required), milestone, +2 more.
Retrieve events using pattern completion. Supports semantic search, temporal filters, participant filters, and effect filters.
Inputs: limit, place, entities, effects_on, +4 more.
Create an episodic Event with Who/Why/What/Where/Effects blocks. Mimics hippocampal CA3 pattern separation and binding.
Inputs: event (required).
Create agent-relative Reflection about an event. Mimics prefrontal cortex self-referential processing with egocentric perspective.
Inputs: agent_id (required), event_id (required), reflection (required).
Execute arbitrary Cypher write operations for creating/updating nodes, relationships, and properties. Use for operations beyond specialized tools.
Inputs: cypher (required), params.
Execute read-only Cypher queries for complex data retrieval. Use for custom queries, analytics, and exploring the graph structure.
Inputs: cypher (required), params.
Retrieve memories using semantic similarity, emotional valence, temporal range, or entity involvement. Returns ranked results combining vector similarity with Cypher filters.
Inputs: limit, query, event_types, temporal_range, +4 more.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 8 tools
There is significant overlap between hippocampus_write_event and encode_memory, both creating episodic memory events with similar semantics. Similarly, hippocampus_search_events and recall_memory both retrieve memories with overlapping filters, making it unclear which tool to select. The distinction between generic graph operations (mutate_graph/query_graph) and specialized tools adds further ambiguity.
Naming mixes a hippocampus_ prefix on three tools (write_event, write_reflection, search_events) with plain verb_noun names on the rest (encode_memory, recall_memory, mutate_graph, evolve_bond, query_graph). While the verbs are fairly descriptive, the inconsistent prefix and overlapping vocabulary (write vs encode, search vs recall) break a predictable pattern.
Eight tools is a reasonable count for a memory and knowledge graph server. However, the presence of redundant event creation and retrieval tools means the count is slightly padded; the true functional set is smaller. Still, the overall number is appropriate for the domain.
The server covers creation (events, reflections), retrieval (search, recall), relationship manipulation (evolve_bond), and custom graph operations (mutate_graph, query_graph). However, there are no dedicated tools for updating or deleting memories, and mutating via raw Cypher is an indirect workaround. Notable gaps exist for lifecycle management of memories.
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Graph-native persistent memory for AI agents — 33 MCP tools, zero-LLM writes.
Persistent, portable memory for AI assistants — your private memory graph, from any MCP client.
Shared cross-LLM long-term memory over MCP: semantic recall, sessions, and media (pgvector).
