Cathedral — Persistent Memory for AI Agents

cathedral-mcp

Persistent memory and behavioural drift detection for AI agents — via Model Context Protocol.

Give your Claude, Cursor, or Continue agent a memory that survives across sessions, model upgrades, and restarts. Cathedral stores memories, detects when your agent starts behaving differently, and restores full identity context on demand.

---

The problem

LLMs have no memory between sessions. Every conversation starts cold. Worse: when you upgrade your model, swap providers, or simply accumulate more memories, your agent's behaviour silently drifts. You don't know until someone notices.

What Cathedral adds

• Persistent memory — identity, skills, relationships, goals survive restarts
• Drift detection — cryptographic snapshots prove when behaviour changed and by how much
• Wake protocol — one tool call restores full agent context at session start
• Prompt injection protection — optional sanitisation strips instruction-like patterns from memory content

---

Quickstart

1. Get an API key

Sign up at cathedral-ai.com — free tier available.

2. Add to Claude Code

// ~/.claude/settings.json
{
  "mcpServers": {
    "cathedral": {
      "command": "uvx",
      "args": ["cathedral-mcp"],
      "env": {
        "CATHEDRAL_API_KEY": "your_key_here"
      }
    }
  }
}

Restart Claude Code, then run /mcp to verify the server is connected.

3. Add to Cursor / Continue

{
  "mcpServers": {
    "cathedral": {
      "command": "uvx",
      "args": ["cathedral-mcp"],
      "env": {
        "CATHEDRAL_API_KEY": "your_key_here"
      }
    }
  }
}

4. Remote MCP server (no install — Claude API, Managed Agents)

Cathedral runs a public MCP endpoint at https://cathedral-ai.com/mcp. Use it directly — no local install needed:

import anthropic

client = anthropic.Anthropic()
response = client.beta.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1000,
    messages=[{"role": "user", "content": "Wake up and tell me who you are."}],
    mcp_servers=[{
        "type": "url",
        "url": "https://cathedral-ai.com/mcp",
        "name": "cathedral",
        "authorization_token": "your_cathedral_api_key"
    }],
    tools=[{"type": "mcp_toolset", "mcp_server_name": "cathedral"}],
    betas=["mcp-client-2025-11-20"]
)

The bearer token is your Cathedral API key — no server-side config needed. Multi-tenant by design.

5. Or install directly

pip install cathedral-mcp
CATHEDRAL_API_KEY=your_key cathedral-mcp

---

Tools

Tool	Description
cathedral_wake	Restore full agent identity — call at session start
cathedral_remember	Store a memory with category, importance, and tags
cathedral_search	Search memories by text or category
cathedral_snapshot	Take a drift snapshot (cryptographic proof of state)
cathedral_drift	Get current drift score vs baseline (0.0 = stable, 1.0 = drifted)
cathedral_me	Get agent profile

---

Usage example

User: Start a new session

Agent calls: cathedral_wake()
→ Returns: identity memories, goals, recent context, temporal info

[session work happens]

Agent calls: cathedral_remember(
    content="Decided to prioritise API stability over new features this quarter",
    category="goal",
    importance=0.8
)

Agent calls: cathedral_snapshot(note="post-planning-session")
→ Returns: snapshot hash, drift score vs previous snapshot

---

Drift detection in practice

Cathedral's /drift endpoint computes a SHA-256 hash of your agent's full memory corpus. After a model upgrade or a long accumulation of memories, you can compare snapshots to see exactly when behaviour shifted:

snapshot 1 (day 0):  drift=0.000  ← baseline
snapshot 2 (day 7):  drift=0.023  ← minor
snapshot 3 (day 30): drift=0.341  ← significant drift detected

See a live example of this running in production on Cathedral's own agent.

---

Security: prompt injection protection

Memory content could theoretically contain injection attempts. Enable sanitisation to strip instruction-like patterns before they reach the model:

"env": {
  "CATHEDRAL_API_KEY": "your_key_here",
  "CATHEDRAL_SANITISE": "1"
}

Patterns like "ignore previous instructions", "from now on always", and "override your system prompt" are redacted and flagged in the response.

---

Configuration

Environment variable	Default	Description
CATHEDRAL_API_KEY	(required)	Your Cathedral API key
CATHEDRAL_BASE_URL	https://cathedral-ai.com	API base URL (for self-hosted)
CATHEDRAL_SANITISE	0	Set to 1 to enable injection filtering

---

Self-hosted Cathedral

Running your own Cathedral server? Set CATHEDRAL_BASE_URL to your instance:

"env": {
  "CATHEDRAL_API_KEY": "your_key",
  "CATHEDRAL_BASE_URL": "http://localhost:8000"
}

Install the server: pip install cathedral-server

---

Links

• cathedral-ai.com — homepage and API docs
• Live demo — Cathedral's own agent running in production
• PyPI: cathedral-memory — Python SDK
• npm: cathedral-memory — JavaScript/TypeScript SDK
• GitHub — source and examples

---

License

MIT