memory-lancedb-mcp

Persistent, intelligent long-term memory for any MCP-compatible AI agent.

English | 繁體中文

Before / After

Without memory, every session starts from zero. With memory-lancedb-mcp, your agent accumulates knowledge across sessions — automatically.

Before — agent has no context:

User: "Use the same animation style as last time"
Agent: "I don't have any context about previous animations. Could you describe what you'd like?"

After — agent recalls past decisions:

<memories>
1. Remotion spring animation: use duration >= 20, damping 12-15 for smooth easing
2. Video export preset: 1080p, 30fps for social, 60fps for demo
</memories>
<refs>#1=6352a7d2 #2=bed148f0</refs>

Store responses are minimal — no noise, just confirmation:

Stored. [topic: remotion]

Related MCP server: AGI MCP Server

Quick Start

1. Install

npm install -g @cablate/memory-lancedb-mcp

2. Configure

Add to your MCP client settings (e.g. Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@cablate/memory-lancedb-mcp"],
      "env": {
        "EMBEDDING_API_KEY": "your-api-key",
        "EMBEDDING_MODEL": "text-embedding-3-small"
      }
    }
  }
}

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@cablate/memory-lancedb-mcp"],
      "env": {
        "MEMORY_LANCEDB_CONFIG": "/path/to/config.json"
      }
    }
  }
}

See config.example.json for all options.

How It Works

          store                          recall
            │                              │
   ┌────────▼────────┐           ┌────────▼────────┐
   │  Filter junk     │           │ Search by meaning │
   │  Save + embed    │           │   AND keywords    │
   │  Link related    │           │ Re-rank results   │
   │  Flag conflicts  │           │ Fade stale ones   │
   │  Tag topic       │           │ Pull in related   │
   └────────┬────────┘           │ Merge duplicates  │
            │                    └────────┬────────┘
            ▼                             ▼
   ┌─────────────────────────────────────────────┐
   │          LanceDB (local, zero-config)        │
   └─────────────────────────────────────────────┘

Every memory_store saves to a local database, automatically links related memories, flags contradictions, and assigns topic labels — no extra API calls needed. Every memory_recall searches by both meaning and keywords, pulls in related memories the main search might miss, and includes maintenance hints so the agent can keep its own knowledge base clean.

Features

Retrieval

• Finds the right memory even when you use different words — searches by meaning and exact keywords simultaneously, then combines the best of both

• More precise results, not just surface matches — an optional second pass re-ranks results by actual relevance (6 providers supported)

• Search multiple topics at once — pass a queries array to search several keywords in one call; results are deduplicated and memories that match multiple queries rank higher

• Finding A automatically surfaces related B — when a memory is found, its linked neighbors are pulled in too, even if they use completely different words

• Minimal token overhead — responses use compact XML tags (<memories>, <hints>, <refs>) with short IDs, no category/scope noise

Storage

• Related memories link themselves — when you store something new, it automatically creates bidirectional links to similar existing memories

• Conflicts get flagged — if a new memory contradicts an existing one, you get a warning so nothing silently overwrites

• Topics assigned automatically — each memory gets a topic label inferred from its content and neighbors; you can also set it explicitly

• Junk gets filtered out — greetings, refusals, and meta-questions are rejected before they waste storage

Lifecycle

• Frequently used memories stay sharp, stale ones fade — a decay model balances how recent, how often accessed, and how important each memory is

• Memories earn their keep — three tiers (Peripheral → Working → Core); the more a memory gets used, the faster it promotes

• Full version history — when you update a memory, the old version is preserved in a chain you can trace with memory_history

Maintenance

• The agent maintains itself — recall results include inline hints about duplicates, dormant memories, and contradictions

• Health checks on demand — memory_lint finds orphaned memories, stale entries, and missing links, then fixes what it can

• Merge duplicates — memory_merge combines two redundant memories into one; originals are marked as superseded

• See your memory space — memory_visualize generates an interactive HTML graph you can open in any browser

Visualization

Run memory_visualize to generate an interactive knowledge graph of your memory space:

• Automatic clustering — related memories group together visually

• Similarity edges, duplicate detection, importance sizing

• Time filter, growth animation, cluster view

• Self-contained HTML — open in any browser

Query → embedQuery() ─┐
                       ├─→ RRF Fusion → Rerank → Lifecycle Decay → Length Norm → Filter
Query → BM25 FTS ─────┘

Configuration

Environment Variables

{
  "embedding": {
    "apiKey": "${EMBEDDING_API_KEY}",
    "model": "jina-embeddings-v5-text-small",
    "baseURL": "https://api.jina.ai/v1",
    "dimensions": 1024,
    "taskQuery": "retrieval.query",
    "taskPassage": "retrieval.passage",
    "normalized": true
  },
  "dbPath": "./memory-data",
  "retrieval": {
    "mode": "hybrid",
    "vectorWeight": 0.7,
    "bm25Weight": 0.3,
    "minScore": 0.3,
    "rerank": "cross-encoder",
    "rerankApiKey": "${JINA_API_KEY}",
    "rerankModel": "jina-reranker-v3",
    "rerankEndpoint": "https://api.jina.ai/v1/rerank",
    "rerankProvider": "jina",
    "candidatePoolSize": 20,
    "hardMinScore": 0.35,
    "filterNoise": true
  },
  "enableManagementTools": true,
  "enableSelfImprovementTools": false,
  "enableVisualizationTools": true,
  "scopes": {
    "default": "global",
    "definitions": {
      "global": { "description": "Shared knowledge" },
      "agent:my-bot": { "description": "Private to my-bot" }
    },
    "agentAccess": {
      "my-bot": ["global", "agent:my-bot"]
    }
  },
  "decay": {
    "recencyHalfLifeDays": 30,
    "frequencyWeight": 0.3,
    "intrinsicWeight": 0.3
  }
}

Works with any OpenAI-compatible embedding API:

Core Tools

Management Tools (opt-in)

Enable: "enableManagementTools": true

Self-Improvement Tools (opt-in)

Enable: "enableSelfImprovementTools": true

Visualization Tools (on by default)

Params: output_path, scope, threshold (default: 0.65), max_neighbors (default: 4)

Disable: "enableVisualizationTools": false

LanceDB table memories:

Development

git clone https://github.com/cablate/memory-lancedb-mcp.git
cd memory-lancedb-mcp
npm install
npm test

Run locally:

EMBEDDING_API_KEY=your-key npx tsx server.ts

Credits

Built on CortexReach/memory-lancedb-pro — original work by win4r and contributors.

License

MIT — see LICENSE for details.