MCP 服务器

Kontext MCP Server

Portable, provider-agnostic memory for AI agents using Azure Data Explorer (Kusto) with temporal and semantic scoring.

README

Kontext MCP Server

Own your Kontext: portable, provider‑agnostic memory for AI agents. Never repeat yourself again.

Kontext transforms Azure Data Explorer (Kusto) into a sophisticated context engine that goes beyond simple vector storage. While traditional vector DBs only store embeddings, Kontext provides layered memory with rich temporal and usage signals—combining recency, frequency, semantic similarity, pins, and decay scoring.

Overview

Kontext provides two powerful MCP tools for intelligent memory management:

`remember`

remember(fact: str, type: str, scope: Optional[str] = "global") -> str

Stores a memory item in the Kusto-backed memory store with automatic embedding generation.

Parameters:

fact: Text to remember
type: Memory type ("fact", "context", or "thought")
scope: Memory scope (defaults to "global")

Returns: Unique ID of the stored memory

`recall`

recall(query: str, filters: Optional[Dict[str, Any]] = None, top_k: int = 10) -> List[Dict[str, Any]]

Retrieves relevant memories using semantic similarity and KQL-powered ranking.

Parameters:

query: Search query for semantic matching
filters: Optional filters (e.g., {"type": "fact", "scope": "global"})
top_k: Maximum number of results to return

Returns: List of memory objects with metadata (id, fact, type, scope, creation_time, sim)

Why Kontext?

The Gap: Agents need intelligent memory that considers not just semantic similarity, but also temporal patterns, usage frequency, and contextual relevance. Most vector databases fall short by ignoring these rich signals and locking you into a single cloud provider.

The Solution: Kontext leverages Kusto's powerful query language (KQL) to score and rank memories using multiple dimensions:

// Conceptual query for scoring memories
Memory 
| extend score = w_t * exp(-ago(ingest)/7d) * 
                 w_f * log(1+hits) * 
                 w_s * cosine_sim * 
                 w_p * pin 
| top 20 by score

Key Benefits

Temporal Reasoning: Native timestamp handling, retention policies, and time-decay scoring
Semantic Retrieval: Built-in vector columns with cosine similarity search
Expressive Ranking: KQL enables complex scoring that weighs time, frequency, pins, and semantics
Cost Effective: Free tier with instant provisioning and predictable scaling
True Portability: Simple MCP API keeps your models and cloud providers interchangeable

Architecture

Agent ⇆ Kontext MCP
         ├── remember(fact, meta)
         └── recall(query, meta)
                  ↓
           Azure Kusto

Ingest: Text splitting → embedding generation → vector + metadata storage
Retrieve: KQL-powered scoring combines temporal, frequency, semantic, and pin signals

Quick Setup

Add Kontext to your MCP settings with the following configuration:

{
  "servers": {
    "kontext": {
      "type": "stdio",
      "command": "uvx",
      "args": ["kontext-mcp"],
      "env": {
        "KUSTO_CLUSTER": "https://your-cluster.kusto.windows.net/",
        "KUSTO_DATABASE": "your-database",
        "KUSTO_TABLE": "Memory",
        "EMBEDDING_URI": "https://your-openai.azure.com/openai/deployments/text-embedding-3-large/embeddings?api-version=2023-05-15;managed_identity=system"
      }
    }
  }
}

Environment Variables:

KUSTO_CLUSTER: Your Azure Data Explorer cluster URL
KUSTO_DATABASE: Database name for storing memories
KUSTO_TABLE: Table name for memory storage (default: "Memory")
EMBEDDING_URI: Azure OpenAI endpoint for embedding generation

Current Features

remember: Store facts with automatic embedding generation using Kusto's ai_embeddings() plugin
recall: Retrieve semantically similar facts using cosine similarity search
FastMCP Integration: Built on the FastMCP framework for easy tool registration and schema generation
Kusto Backend: Leverages Azure Data Explorer for scalable storage and querying

Roadmap

Advanced Scoring: Multi-dimensional ranking with temporal decay, frequency weighting, and pin support
Memory Tiers: Short-term context, working memory, and long-term fact storage
Hosted Embeddings: Optional E5 model hosting to reduce setup friction
Enhanced Caching: Multi-tier memory management and query optimization