memory-lancedb MCP sunucusu

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]

---

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 ─────┘

Stage	Effect
RRF Fusion	Combines semantic and exact-match recall
Cross-Encoder Rerank	Promotes semantically precise hits
Lifecycle Decay	Weibull freshness + access frequency + importance
Length Normalization	Prevents long entries from dominating (anchor: 500 chars)
Hard Min Score	Removes irrelevant results (default: 0.35)
MMR Diversity	Cosine similarity > 0.85 → demoted

---

Configuration

Environment Variables

Variable	Required	Description
EMBEDDING_API_KEY	Yes	API key for embedding provider
EMBEDDING_MODEL	No	Model name (default: text-embedding-3-small)
EMBEDDING_BASE_URL	No	Custom base URL for non-OpenAI providers
MEMORY_DB_PATH	No	LanceDB storage directory
MEMORY_LANCEDB_CONFIG	No	Path to JSON config file

{
  "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:

Provider	Model	Base URL	Dimensions
OpenAI	text-embedding-3-small	https://api.openai.com/v1	1536
Jina	jina-embeddings-v5-text-small	https://api.jina.ai/v1	1024
DeepInfra	Qwen/Qwen3-Embedding-8B	https://api.deepinfra.com/v1/openai	1024
Google Gemini	gemini-embedding-001	https://generativelanguage.googleapis.com/v1beta/openai/	3072
Ollama (local)	nomic-embed-text	http://localhost:11434/v1	varies

Provider	rerankProvider	Endpoint	Example Model
Jina	jina	https://api.jina.ai/v1/rerank	jina-reranker-v3
Hugging Face TEI	tei	http://host:8081/rerank	BAAI/bge-reranker-v2-m3
SiliconFlow	siliconflow	https://api.siliconflow.com/v1/rerank	BAAI/bge-reranker-v2-m3
Voyage AI	voyage	https://api.voyageai.com/v1/rerank	rerank-2.5
Pinecone	pinecone	https://api.pinecone.io/rerank	bge-reranker-v2-m3
DashScope	dashscope	https://dashscope.aliyuncs.com/api/v1/services/rerank	gte-rerank

---

Core Tools

Tool	Description
memory_recall	Search memories — supports batch queries, relation expansion, topic filtering, and inline maintenance hints
memory_store	Save a memory — auto-links related ones, flags contradictions, infers topic, filters junk
memory_forget	Delete by ID or search query
memory_update	Update a memory; the old version is preserved in a version chain
memory_merge	Merge two memories into one
memory_history	Trace version history through update/merge chains

Management Tools (opt-in)

Tool	Description
memory_stats	Usage statistics by scope and category
memory_list	List recent memories with filtering
memory_lint	Health checks + auto-fix missing relations

Enable: "enableManagementTools": true

Self-Improvement Tools (opt-in)

Tool	Description
self_improvement_log	Log structured learning/error entries
self_improvement_extract_skill	Create skill scaffolds from learnings
self_improvement_review	Summarize governance backlog

Enable: "enableSelfImprovementTools": true

Visualization Tools (on by default)

Tool	Description
memory_visualize	Generate interactive HTML memory graph

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

Disable: "enableVisualizationTools": false

---

LanceDB table memories:

Field	Type	Description
id	string (UUID)	Primary key
text	string	Memory text (FTS indexed)
vector	float[]	Embedding vector
category	string	preference / fact / decision / entity / skill / lesson / other
scope	string	Scope identifier
importance	float	Importance score 0-1
timestamp	int64	Creation timestamp (ms)
metadata	string (JSON)	Extended metadata (tier, access_count, relations, topic, etc.)

---

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.