🧠 Sovereign Memory Daemon

A shared memory layer for AI coding agents. Stop your team's 10 Claude instances from writing the same code 10 different ways.

Runs 100% locally. SQLite + ONNX embeddings. No cloud. No telemetry. Zero outbound connections.

What It Does

Every AI coding tool (Claude, Cursor, Copilot) operates in isolation. Each one has zero memory of what the others have built. This daemon solves that by providing a persistent, shared brain that every AI agent queries before writing code.

• Ingests your entire codebase via Tree-Sitter AST parsing
• Stores semantic embeddings in a local SQLite + HNSW vector index
• Serves memory to any MCP-compatible AI tool (Claude Desktop, Cursor, Zed)
• Scales to teams via SSE transport + API key auth + Prometheus observability

Quick Start (Solo Developer)

# Option A: npx (zero-install)
npx sovereign-brain

# Option B: Clone
git clone <repo> && cd sovereign-brain && npm install
npm run install-claude  # Auto-configure Claude Desktop
npm start               # Boots daemon (auto-ingests on first run)

Done. Open Claude Desktop → your brain is live with 5 MCP tools:

• search_memory — Hybrid semantic + keyword search
• add_memory — Store knowledge across sessions (with conflict detection)
• build_context — Generate LLM-optimized context packs
• get_stats — Brain health metrics
• find_duplicates — Detect code duplication

Team Deployment (Shared Brain)

Host the daemon on a central server so every developer's AI agent talks to the same brain.

1. Server Setup

# Generate team API keys (one per developer)
export MEMORY_API_KEYS="key_alice_abc123,key_bob_def456,key_carol_ghi789"

# Boot in team mode (binds to 0.0.0.0 instead of localhost)
npm start

2. Developer Setup (Remote Claude Desktop)

Each developer adds this to their claude_desktop_config.json:

{
  "mcpServers": {
    "team-brain": {
      "url": "http://team-brain.internal:31337/mcp/sse?token=key_alice_abc123"
    }
  }
}

3. Observability

Prometheus metrics are exposed at /metrics:

brain_memories_added_total
brain_searches_total
brain_search_duration_seconds
process_cpu_seconds_total
nodejs_heap_size_total_bytes

Scrape with Datadog, Grafana, or any Prometheus-compatible system.

Architecture

┌──────────────────────────────────────────────┐
│              Sovereign Brain Daemon          │
│                                              │
│  ┌──────────┐  ┌───────────┐  ┌───────────┐ │
│  │ SQLite   │  │ HNSW      │  │ Tree-     │ │
│  │ FTS5+WAL │  │ Vector    │  │ Sitter    │ │
│  │          │  │ Index     │  │ AST       │ │
│  └──────────┘  └───────────┘  └───────────┘ │
│                                              │
│  ┌──────────┐  ┌───────────┐  ┌───────────┐ │
│  │ MCP      │  │ REST API  │  │ Prometheus│ │
│  │ SSE/stdio│  │ :31337    │  │ /metrics  │ │
│  └──────────┘  └───────────┘  └───────────┘ │
└──────────────────────────────────────────────┘
        ↑               ↑              ↑
   Claude Desktop    VS Code        Grafana
   Cursor/Zed       Extension       Datadog

Environment Variables

| Variable | Default | Description | |---|---|---| | MEMORY_BRAIN_DB_PATH | .memory-brain/brain.db | SQLite database location | | MEMORY_API_KEYS | (none) | Comma-separated team API keys | | MEMORY_TEAM_MODE | false | Force team mode (0.0.0.0 binding) |

Conflict Detection

When any agent stores a memory via add_memory, the daemon scans memories added in the last hour for high-similarity entries (>85% cosine similarity). If found, the MCP response includes a conflict warning:

Memory stored (id: 42, dimension: code)

⚠️ CONFLICT WARNING ⚠️
High similarity to recently added memories (potential contradiction):
- [Similarity: 0.91] Auth module uses session cookies for state management
  Resolution: Review both memories — high similarity suggests potential duplication or contradiction.

This catches contradictory architectural decisions before they ship — e.g. when Developer A's Claude decides "use JWT" while Developer B's Claude decides "use session cookies."

CLI

npm run memory -- add --type code --content "Auth uses JWT" --project myapp
npm run memory -- search --query "authentication" --format json
npm run memory -- ingest-repo ./
npm run memory -- context --query "payment flow" --target claude

License

MIT. No cloud. No spies. Ship it.