ContextBridge

A zero-dependency MCP memory server for Windows that gives AI coding assistants a shared, persistent memory layer with semantic search.

What is ContextBridge?

AI coding assistants like Claude Code are stateless—each session starts fresh, with no persistent memory of your prior context, decisions, or discoveries. ContextBridge solves this by running as a background Windows Service that all your AI tools connect to. It stores memories persistently, searchable by meaning (not keywords), so context accumulates across sessions and across different tools simultaneously.

Why ContextBridge?

• Zero external dependencies — No Docker, Postgres, Python, or Ollama. Everything runs in one Windows Service process.
• In-process embeddings — Embedding model (all-MiniLM-L6-v2, ~22 MB) runs via ONNX Runtime; no external API calls.
• Single-file storage — SQLite database at %ProgramData%\ContextBridge\memories.db. Backup is one file copy.
• Shared across tools — Claude Code (HTTP), Claude Desktop (stdio), Cline, VS Code Chat Agents all connect to the same memory store in real-time.
• Memory survives service restarts — Memories persist in SQLite; the embedding model loads once on startup and stays warm.

Requirements

• Windows 10 / Windows 11 (required for Windows Service integration)
• Option A: .NET 10 SDK (if installing via dotnet tool)
• Option B: Windows 10+ (no additional prerequisites if downloading a pre-built .exe)
• Admin PowerShell (required to install/uninstall the Windows Service)

Quick Start

Option A: Install via dotnet global tool

If you have the .NET 10 SDK installed:

dotnet tool install -g ThunderEagle.ContextBridge

This places a context-bridge executable on your system PATH.

Option B: Direct download

Download a pre-built executable from GitHub Releases (no SDK required; one-time SmartScreen warning). Add the .exe to your PATH or invoke it directly.

First-run setup

Run these two commands in admin PowerShell (right-click → "Run as Administrator"):

# 1. Download the embedding model, register as Windows Service, start it
context-bridge service install

# 2. Configure your AI tools to connect
context-bridge configure

What happens:

1. service install downloads the embedding model (~22 MB), registers the Windows Service, sets it to auto-start on boot, and starts it immediately.
2. configure auto-detects installed clients (Claude Code, Claude Desktop, Cline, VS Code Chat) and wires them up to connect to the service.

The service now runs in the background. Your AI tools will see the memory store the next time you restart them.

Verify the service is running:

Get-Service ContextBridge

Expected output: Status = Running

Supported Clients

All clients share the same SQLite database via concurrent connections; memories written by one tool are immediately visible to others.

MCP Tools

ContextBridge exposes seven MCP tools for your AI assistants to use:

Tag conventions (optional; assigned by the AI tool):

• project:<repo-name> — scope memories to a project
• type:decision — architectural or technology choices
• type:preference — coding style, tooling, workflow preferences
• type:pattern — recurring patterns or conventions
• type:reference — pointers to external resources or documentation

Handoff — Resuming Sessions

Memories are permanent facts. A handoff is something different: ephemeral session state that lets you resume where you left off in a future session, without reloading conversation history.

How it works

At the end of a session, ask your AI assistant to save its state:

"Save a handoff for project context-bridge with what we were working on."

The model calls handoff_write with a summary of current work — decisions made, next steps, open questions — scoped to the project.

At the start of the next session, the model calls handoff_list automatically (via server instructions) and incorporates any prior handoff as its opening context. It then calls handoff_acknowledge to remove the handoff once processed.

Handoff tools

Key parameters for handoff_write:

• content — the session summary (free-form text)
• project — project identifier, e.g. context-bridge (optional but recommended)
• ttl_days — how many days to keep the handoff before auto-expiry (default: 7)

Handoffs vs. memories

Do not convert handoff content to memories automatically. Memories are for facts that will remain true indefinitely. If something from a resumed session rises to that level, write it via memory_write separately.

Explicit resumption

If your MCP client supports the prompts capability (e.g. Claude Code), you can trigger a session resumption explicitly:

/mcp__context-bridge__resume-session context-bridge

This invokes the resume-session named prompt, which tells the model to look up any handoff for the specified project and incorporate it.

Expiry and reliability

Handoffs expire after ttl_days and are purged on service startup. If a session crashes before handoff_acknowledge is called, the handoff survives until its TTL — it will surface again in the next session's handoff_list call.

Importing Existing Context

If you've been using Claude Code's built-in file-based memory (~/.claude/projects/<name>/memory/), you can migrate that context into ContextBridge without any special tooling. Just ask:

"Check your memory files and add any relevant entries to context-bridge using memory_batch_write."

Claude Code reads its own memory index, iterates the entries, and calls memory_batch_write to store them in ContextBridge — where they become semantically searchable and visible to all connected clients immediately.

You can scope the request: "import only entries tagged project:my-repo" or "add everything in your memory files."

Once imported, you can remove the original file-based entries to avoid maintaining two stores. ContextBridge becomes the single source of truth, shared across Claude Code, Claude Desktop, and any other connected client.

CLI Reference

All functionality is controlled via the context-bridge command. Run from any command-line (admin PowerShell required for service install/uninstall/config set).

Service Management

context-bridge service install      # Download model, register service, start it
context-bridge service start        # Start the service (no-op if already running)
context-bridge service stop         # Stop the service gracefully
context-bridge service status       # Show current status
context-bridge service uninstall    # Stop and unregister (preserves memories.db)

Model Management

context-bridge model download       # Download embedding model (~22 MB) from Hugging Face
context-bridge model download --yes # Skip confirmation, force re-download

Configuration

context-bridge config get port      # Show current HTTP port (default: 5290)
context-bridge config set port 8000 # Change HTTP port (requires service restart)

Configuration is stored in %ProgramData%\ContextBridge\appsettings.json. Changes take effect on next service start.

Client Configuration

context-bridge configure            # Auto-configure all installed clients

This command:

• Detects installed AI clients (Claude Code, Claude Desktop, Cline, VS Code)
• Writes MCP server configuration for each client
• For Claude Code: injects usage guidelines into ~/.claude/CLAUDE.md
• Prints which clients were configured

Example output:

Claude Code configured (HTTP transport)
Claude Desktop configured (stdio transport)
Cline configured (HTTP transport)
VS Code Chat Agents configured (HTTP transport)

Configured 4 client(s). Restart them to pick up changes.

Run this again after:

• Installing a new AI client
• Changing the service port (config set port)
• Updating ContextBridge itself

Security

Security model: Kestrel binds exclusively to 127.0.0.1 (localhost only). No authentication, no TLS.