Hebbrix MCP Server

A Model Context Protocol server that gives any AI agent long-term memory and a temporal knowledge graph, backed by Hebbrix.

Your agent forgets everything when the session ends. This fixes that, and goes further than a plain memory store:

• Memory — store, search, correct, and version facts across sessions

• Knowledge graph — entities, relationships, timelines, and "what was true at time X"

• Reasoning — ask how confident the agent should be before acting, and log outcomes so it improves

Works with Claude Desktop, Claude Code, Cursor, Cline, Continue, and any other MCP client.

Quick start (no account needed)

Add this to your MCP client config. On first run with no API key, the server mints a free agent account automatically (no email, no dashboard, ~1 second) and saves it to ~/.hebbrix/config.json.

{
  "mcpServers": {
    "hebbrix": { "command": "uvx", "args": ["hebbrix-mcp"] }
  }
}

Restart the client. Done — your agent now has persistent memory.

The free agent account includes 300 learning events and 2,000 retrievals, and expires 14 days after last use if unclaimed. Every tool result carries a hebbrix_usage block so the agent always knows where it stands and will tell you when it's time to claim.

Keep it forever (same key, all memories carry over, unlocks the free monthly tier):

uvx hebbrix-mcp claim --email you@example.com

Related MCP server: Bi-Temporal Knowledge Graph MCP Server

Configuration

Get an API key at hebbrix.com/dashboard/api-keys to use your own account instead of agent mode.

{
  "mcpServers": {
    "hebbrix": {
      "command": "uvx",
      "args": ["hebbrix-mcp"],
      "env": {
        "HEBBRIX_API_KEY": "mem_sk_...",
        "HEBBRIX_COLLECTION_ID": "your-default-collection-uuid"
      }
    }
  }
}

claude mcp add hebbrix -- uvx hebbrix-mcp

{
  "mcpServers": {
    "hebbrix": { "command": "uvx", "args": ["hebbrix-mcp"] }
  }
}

Point your MCP servers config at the uvx hebbrix-mcp command (stdio). Same shape as above. Set HEBBRIX_API_KEY in env to skip agent mode.

The env var always wins over saved agent-mode credentials.

Environment variables

All optional. With nothing set, the server starts in agent mode.

Available Tools

A server-level instruction block teaches the model when to reach for each tool, so a well-behaved agent searches before answering and remembers what matters without being told.

Memory

• hebbrix_remember - Store a fact, decision, or preference.

  • content (string, required): the memory text

  • tags (list, optional), collection_id (string, optional)

  • verbatim (bool, optional): store exactly as given, skip fact-extraction

• hebbrix_search - Semantic search (hybrid vector + BM25 + graph retrieval).

  • query (string, required), limit (int, optional), collection_id (string, optional)

• hebbrix_get - Fetch one memory by id, with metadata.

• hebbrix_update - Correct a memory in place (old versions are kept).

• hebbrix_forget - Delete a memory by id.

• hebbrix_list - List recent memories.

• hebbrix_history - See how a memory changed over time.

Knowledge graph — reads work on every tier (including agent mode); graph writes/inference need a Pro plan.

• hebbrix_search_entities - List known entities (people, orgs, tools, places).

• hebbrix_entity_timeline - What was true about an entity, and when.

• hebbrix_graph_query - Query relationships; pass a timestamp to ask about a point in time.

• hebbrix_contradictions - Surface facts that conflict with each other.

Reasoning & account

• hebbrix_confidence - How confident should the agent be before acting? Grounded in memory + past outcomes.

• hebbrix_log_decision - Record a decision and its outcome; feeds future confidence.

• hebbrix_list_collections - List the memory spaces this key can use.

• hebbrix_account_status - Tier, usage, limits, and expiry.

The server also exposes a hebbrix://profile resource and a context prompt that inject the user's compiled profile.

Running modes

Local (default) — stdio. What the quick start does: one process per client.

Self-hosted HTTP — one instance, your machines:

HEBBRIX_API_KEY=mem_sk_... uvx hebbrix-mcp --transport streamable-http
# serves http://127.0.0.1:8080/mcp

Hosted multi-tenant — one instance, many users. The server holds no key; every request authenticates with its own Authorization header:

HEBBRIX_MCP_MULTI_TENANT=1 HEBBRIX_MCP_HOST=0.0.0.0 uvx hebbrix-mcp --transport streamable-http

{ "mcpServers": { "hebbrix": {
  "url": "https://your-host/mcp",
  "headers": { "Authorization": "Bearer mem_sk_..." }
}}}

In multi-tenant mode there is no default collection — pass collection_id on tool calls.

How it works

┌──────────────────┐   MCP (stdio or HTTP)   ┌─────────────┐    HTTPS     ┌──────────┐
│ Claude / Cursor / │ ───────────────────────→│ hebbrix-mcp │─────────────→│ Hebbrix  │
│ Cline / any agent │      tool calls         │   (this)    │   REST API   │  cloud   │
└──────────────────┘                          └─────────────┘              └──────────┘

This package owns zero state. Tool calls become REST calls against your Hebbrix account; memories, embeddings, the knowledge graph, and retrieval all live in the Hebbrix backend. Delete this package and your memories are still there.

Agent-mode accounts never break mid-task: when a limit is reached you get a structured error with a resolve field, not a failure. Writes stop before reads; reads keep working; the account goes read-only before it expires.

Debugging

Inspect the server with the MCP Inspector:

npx @modelcontextprotocol/inspector uvx hebbrix-mcp

Common issues:

• HTTP 401 on every call — the key is wrong or revoked. Unset HEBBRIX_API_KEY, delete ~/.hebbrix/config.json, and restart to re-provision, or paste a fresh key from the dashboard.

• Agent mode won't start (auto-signup unavailable) — signup may be at daily capacity or your network blocks the API. Set HEBBRIX_API_KEY instead.

• claim says EMAIL_IN_USE — claiming needs an email with no existing Hebbrix account. Use a fresh address (a you+agent@gmail.com alias works).

• A memory isn't searchable immediately — indexing is asynchronous; typical convergence is under 30 seconds.

Development

git clone https://github.com/Hebbrix/hebbrix-mcp
cd hebbrix-mcp
./quick_setup.sh            # venv + editable install
source venv/bin/activate
pytest tests/ -q            # 11 offline tests, no network needed
hebbrix-mcp                 # starts in agent mode on stdio

See CONTRIBUTING.md and CHANGELOG.md.

License

MIT — see LICENSE.

Links

• Hebbrix documentation

• MCP integration guide

• Model Context Protocol