sequential-thinking-mcp-v2

Overview

Sequential Thinking MCP v2 is a local MCP server that implements structured reasoning, session persistence, and project memory management. It follows a first principle thinking framework with an approved design for a layered memory architecture. The server manages reasoning lifecycles through scratch reasoning, checkpoint summaries, promoted conclusions, and canonical docs. It plans to integrate cloud-assisted operations with local-first fallback when network access fails.

Try asking AI

After installing, here are 7 things you can ask your AI assistant:

• retrieve_context

  Retrieve relevant context for the current task

  Try:Retrieve relevant context for this coding taskTry:Find context related to authentication
• search_docs

  Search through canonical documentation

  Try:Search for API documentationTry:Find design documents related to authentication
• search_memory

  Search through promoted memory artifacts

  Try:Search previous conclusions about performance issuesTry:Find memory related to database optimization
• promote_memory

  Promote reasoning to durable memory

  Try:Promote this conclusion to memory for future reuseTry:Save this investigation summary as a memory artifact
• summarize_session

  Create a summary of the current session

  Try:Summarize this session's progressTry:Create a checkpoint summary of the work done
• generate_handoff

  Generate a handoff bundle for session continuity

  Try:Create a handoff document for this sessionTry:Generate a summary for another agent to continue this work
• map_changed_modules

  Map modules that have changed during the session

  Try:Show me which modules have been modifiedTry:Map changes to the codebase since last checkpoint

Comparable tools

Installation

git clone https://github.com/spideynolove/sequential-thinking-mcp-v2.git
cd sequential-thinking-mcp-v2

uv sync
uv run main.py

{
  "mcpServers": {
    "sequential-thinking": {
      "command": "python",
      "args": ["main.py"],
      "cwd": "/path/to/sequential-thinking-mcp-v2"
    }
  }
}

FAQ

What is the memory architecture design?

The design uses markdown as canonical human truth, qmd for document retrieval, Zvec for promoted artifacts, and a structured metadata store for lifecycle state.

How does token management work?

The project targets an active context ceiling of 64K with compaction triggers at 45K. It retrieves on demand, checkpoints at subtask boundaries, and compresses old context.

What happens when network access fails?

The server operates with a cloud-assisted, local-first fallback approach, ensuring functionality even when internet access is unavailable.

Compare sequential-thinking-mcp-v2 with