MemoryExchange

A .NET tool that makes markdown knowledge files (such as Cline-style "Memory Banks") searchable by GitHub Copilot (or any MCP-compatible client) via the Model Context Protocol (MCP).

It pre-processes markdown files into semantically chunked, vector-embedded documents and serves them through an MCP server with tools for searching, retrieving files, and diagnosing index health.

Features

• Fully local by default — SQLite + ONNX (all-MiniLM-L6-v2) embeddings, no cloud dependencies
• Azure provider available — Azure AI Search + Azure OpenAI for cloud-scale deployments
• Hybrid search — FTS5 BM25 keyword search + cosine vector similarity, merged via Reciprocal Rank Fusion (RRF)
• Domain-aware boosting — chunks matching the user's current file domain are boosted (1.3x)
• Instruction boosting — .instructions.md files get additional priority (1.2x)
• Heading-aware chunking — markdown is split by headings with breadcrumb context; code blocks are kept atomic
• Zero-install via dotnet dnx — run directly with dotnet dnx --yes MemoryExchange
• Also installable as a .NET global tool — dotnet tool install -g MemoryExchange

Quick Start

Configure your MCP client

Add the server to your MCP client configuration (e.g. VS Code mcp.json, Cline, etc.):

Option A: Via dotnet dnx (recommended — no install required, .NET 10+)

{
  "servers": {
    "memory-exchange": {
      "type": "stdio",
      "command": "dotnet",
      "args": [
        "dnx",
        "--yes",
        "MemoryExchange",
        "--source-path", "/path/to/your/memory-exchange",
        "--provider", "local",
        "--build-index"
      ]
    }
  }
}

Option B: As a .NET global tool

dotnet tool install -g MemoryExchange

{
  "servers": {
    "memory-exchange": {
      "type": "stdio",
      "command": "memory-exchange",
      "args": [
        "--source-path", "/path/to/your/memory-exchange",
        "--provider", "local",
        "--build-index"
      ]
    }
  }
}

Option C: Environment variables

{
  "servers": {
    "memory-exchange": {
      "type": "stdio",
      "command": "dotnet",
      "args": [
        "dnx",
        "--yes",
        "MemoryExchange",
        "--build-index"
      ],
      "env": {
        "MEMORYEXCHANGE_SOURCEPATH": "/path/to/your/memory-exchange",
        "MEMORYEXCHANGE_PROVIDER": "local"
      }
    }
  }
}

Option D: From source

{
  "servers": {
    "memory-exchange": {
      "type": "stdio",
      "command": "dotnet",
      "args": [
        "run",
        "--project", "/path/to/MemoryExchange/src/MemoryExchange.McpServer",
        "--",
        "--source-path", "/path/to/your/memory-exchange",
        "--provider", "local",
        "--build-index"
      ]
    }
  }
}

With --build-index, the server indexes your source directory on startup (incremental — only changed files are re-processed) and then starts accepting MCP connections. No separate indexer step required.

Alternatively, use --watch instead of --build-index to keep the index updated automatically while the server runs. The watcher builds the index on startup and then monitors for file changes (creates, edits, deletes, renames), triggering incremental re-indexing as needed.

Stand-alone Indexer

You can also index separately using the CLI indexer:

dotnet run --project src/MemoryExchange.Indexer -- \
  --source /path/to/your/memory-exchange \
  --provider local \
  --database-path /path/to/memoryexchange.db

Flag	Description
--source, -s	(required) Path to the source directory containing markdown files
--provider, -p	local (default) or azure
--database-path	SQLite database file path (defaults to memory_exchange.db in the source directory)
--model-path	Custom ONNX model file path (local provider)
--index-name	Logical index name (default: memory-exchange)
--force, -f	Force full re-index, ignoring cached state
--exclude	Glob patterns to exclude from indexing (can be specified multiple times)
--watch, -w	After indexing, watch for file changes and re-index incrementally (runs until Ctrl+C)

Configuration

Configuration is loaded with the following precedence (highest wins):

1. appsettings.json
2. Environment variables (MEMORYEXCHANGE_ prefix)
3. CLI arguments

Environment Variables

Variable	Maps to
MEMORYEXCHANGE_SOURCEPATH	Source directory path
MEMORYEXCHANGE_PROVIDER	local or azure
MEMORYEXCHANGE_INDEXNAME	Search index name
MEMORYEXCHANGE_DATABASEPATH	SQLite database path (defaults to memory_exchange.db in the source directory)
MEMORYEXCHANGE_MODELPATH	ONNX model file path
MEMORYEXCHANGE_BUILDINDEX	Set to true to build index on startup
MEMORYEXCHANGE_WATCH	Set to true to enable file watching mode
MEMORYEXCHANGE_DEBUG	Set to true to enable debug-only tools (e.g. get_index_status)
MEMORYEXCHANGE_AZURE_SEARCH_ENDPOINT	Azure AI Search endpoint
MEMORYEXCHANGE_AZURE_SEARCH_APIKEY	Azure AI Search API key
MEMORYEXCHANGE_AZURE_OPENAI_ENDPOINT	Azure OpenAI endpoint
MEMORYEXCHANGE_AZURE_OPENAI_APIKEY	Azure OpenAI API key
MEMORYEXCHANGE_AZURE_OPENAI_DEPLOYMENT	Azure OpenAI embedding deployment name

MCP Server CLI Args

--source-path, --provider, --index-name, --database-path, --model-path, --build-index, --watch, --debug, --exclude, --azure-search-endpoint, --azure-search-key, --azure-openai-endpoint, --azure-openai-key, --azure-openai-deployment

--build-index — When present, the MCP server runs incremental indexing on startup before accepting connections. This eliminates the need to run the Indexer CLI separately. Requires --source-path to be set.

--watch — Monitors the source directory for markdown file changes and triggers incremental re-indexing automatically. Implies --build-index on startup (builds the index if it doesn't exist, then watches for changes). When --watch is active, a separate --build-index flag is redundant. Requires --source-path to be set.

--debug — Enables debug-only MCP tools (e.g. get_index_status). Without this flag, these tools are not registered and are invisible to AI agents.

--exclude — Optional glob patterns to exclude files/directories from indexing. Can be specified multiple times (e.g., --exclude "**/archive/**" --exclude "**/drafts/**"). The personal/ directory is always excluded regardless of this setting.

--database-path — If omitted, defaults to memory_exchange.db inside the --source-path directory.

MCP Tools

The server exposes the following tools:

search_memory_bank

Parameter	Type	Required	Description
query	string	yes	What to search for (e.g. "how does caching work", "deployment architecture")
currentFilePath	string	no	File the user is editing — enables domain-aware boosting
topK	int	no	Number of results, 1-10 (default: 5)

get_memory_file

Retrieves the full content of a specific markdown file from the memory exchange. Useful when a search result references a file and the agent needs the complete context.

Parameter	Type	Required	Description
filePath	string	yes	Relative path to the markdown file (e.g. "architecture/database.md")

get_index_status (debug only)

Returns diagnostic information about the memory exchange index. Use this to verify the index is populated before searching, or to diagnose why search_memory_bank returns no results.

No parameters. Returns the source path, provider, index name, chunk count, file count, and last indexed time. When the index is empty, it includes actionable troubleshooting hints.

This tool is only registered when the server is started with the --debug flag (or MEMORYEXCHANGE_DEBUG=true). Without it, AI agents will not see this tool.

Startup Diagnostics

When the MCP server starts, it prints diagnostic information to stderr including the configured provider, source path, index name, watch/build-index mode, and the number of markdown files found in the source directory. This output is visible in the MCP client's server logs and is useful for confirming the server is configured correctly.

Architecture

src/
  MemoryExchange.Core/          Core abstractions, chunking, search orchestrator
  MemoryExchange.Azure/         Azure AI Search + Azure OpenAI provider
  MemoryExchange.Local/         SQLite + ONNX provider (default)
  MemoryExchange.Indexing/      Shared indexing pipeline (FileScanner + IndexingPipeline)
  MemoryExchange.Indexer/       CLI indexer tool
  MemoryExchange.McpServer/     MCP stdio server (supports --build-index)
tests/
  MemoryExchange.Indexer.LocalRunner/   Development test harness

Provider Abstractions

• IEmbeddingService — generate vector embeddings from text
• ISearchService — read-side: hybrid search (keyword + vector)
• ISearchIndex — write-side: upsert/delete chunks

Search Pipeline

1. Query is embedded via IEmbeddingService
2. ISearchService runs FTS5 BM25 + vector cosine similarity, merged via RRF (k=60)
3. SearchOrchestrator applies domain boosting (1.3x) and instruction boosting (1.2x)
4. Results are over-fetched at 2x, reranked, and trimmed to topK

Local Provider Details

• Model: all-MiniLM-L6-v2 (384-dim, L2-normalized)
• Tokenizer: Custom WordPieceTokenizer with embedded 30k vocab
• Storage: SQLite with WAL mode, FTS5 virtual table, embeddings as BLOBs
• Vector search: Pure C# cosine similarity (dot product on normalized vectors)

Building from Source

Requires .NET 9.0 SDK.

1. Download the ONNX model

The published NuGet package bundles the model, but when building from source you need to download it manually. Use the included script to fetch the all-MiniLM-L6-v2 embedding model (~90 MB) from HuggingFace:

# Windows (PowerShell)
.\scripts\download-model.ps1

# Linux / macOS
./scripts/download-model.sh

This places the model in src/MemoryExchange.Local/Models/all-MiniLM-L6-v2.onnx. You can pass a custom output path as an argument if needed.

2. Build

dotnet build MemoryExchange.slnx

License

MIT - Copyright (c) 2026 Moaid Hathot