Sequential Memory MCP Server

A Model Context Protocol (MCP) server that combines sequential thinking with persistent memory through a knowledge graph. This enables AI assistants to explore decision trees by recording thinking traces, branching at low-confidence points, and backtracking to explore alternative paths.

Features

• Linear thinking when confidence is high (≥ 0.6)

• Automatic branching when confidence drops below threshold (< 0.6)

• Path recording in an in-memory graph structure

• Backtracking to explore alternative branches

• Path visualization to see complete thinking traces

• Unexplored branch detection to identify paths not yet taken

Related MCP server: memory-bank-mcp

Installation

1. Clone this repository

2. Create a virtual environment:

   python -m venv venv
   venv\Scripts\activate  # On Windows
   # or
   source venv/bin/activate  # On Unix/MacOS

3. Install dependencies:

   pip install mcp

Configuration

Add the following to your Claude Desktop configuration file: C:\Users\steve\AppData\Roaming\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "sequential-memory": {
      "command": "python",
      "args": [
        "-m",
        "sequential_memory.server"
      ],
      "cwd": "C:\\Users\\steve\\claude\\seq-mem\\src"
    }
  }
}

Usage

The server provides 5 main tools:

1. think

Process a thought with a confidence level.

• Parameters:

  • thought (string): The thought content

  • confidence (number): Confidence level (0.0-1.0)

• Returns: Status (continue/branch), current node ID, and whether alternatives are needed

2. select_path

Choose from alternative thoughts at a branch point.

• Parameters:

  • alternatives (array): List of alternative thoughts with confidence levels

  • selected_index (integer): Which alternative to select (0-based)

• Returns: Selected thought information and new current node ID

3. backtrack

Return to the last high-confidence node in the current path.

• Parameters: None

• Returns: Information about the node backtracked to (or no_target if none found)

4. show_current_path

Display the current thinking path from root to current node.

• Parameters: None

• Returns: Complete path with node details, total nodes, and branch points

5. get_unexplored_branches

Find all branch points with unexplored alternatives.

• Parameters: None

• Returns: List of unexplored branches with their alternatives

Example Usage

User: Let's think through a problem step by step.

Claude: I'll use sequential thinking to explore this systematically.

[thinks: "First, let me understand the problem clearly" (0.8)]
[thinks: "The key aspects seem to be X, Y, and Z" (0.9)]
[thinks: "Now I need to decide on an approach..." (0.3)]

I've reached a point of uncertainty. Let me explore different approaches:

[select_path with alternatives:
  - "Focus on aspect X first" (0.7)
  - "Start with aspect Y" (0.6)
  - "Consider aspect Z" (0.5)]

[Continue thinking on selected path...]

User: Let's backtrack and try a different approach.

Claude: [backtrack]
I've returned to "The key aspects seem to be X, Y, and Z". 
Let me explore one of the other approaches...

Testing

Run the test suite:

python -m pytest tests/test_basic.py
# or
python tests/test_basic.py

Architecture

• graph.py: Core graph data structures (Node, Edge, ThoughtGraph)

• tools.py: MCP tool implementations and definitions

• server.py: Main MCP server implementation

• test_basic.py: Comprehensive test suite

Development

The codebase is organized for clarity and extensibility:

• All graph operations are encapsulated in the ThoughtGraph class

• Tool handlers are separated in the SequentialMemoryTools class

• The server handles only MCP protocol communication

• Tests cover both unit and integration scenarios