Loom-MCP

Loom-MCP is a reasoning-management engine derived from the Microsoft Memento research architecture. It facilitates the generation of high-quality reasoning traces and allows for manual context-window optimization in agentic workflows.

An MCP server for Recursive Reasoning Compaction and SFT data generation. Built with the TypeScript MCP SDK and SQLite.

---

Overview

Current LLMs suffer from "Reasoning Drift" and context saturation during complex, multi-step tasks. Loom-MCP solves this by providing a structured framework for compacting reasoning. It allows an LLM to:

1. Offload raw reasoning ("The Weft") into persistent "Blocks"
2. Compress each block into a concise summary ("The Warp")
3. Export the entire reasoning trajectory as a Memento-spec SFT (Supervised Fine-Tuning) dataset

The exported reasoning traces follow the structural markers from the Microsoft Memento research paper:

<think>
<|block_start|>…raw reasoning…<|block_end|>
<|summary_start|>…concise summary…<|summary_end|>
</think>

---

Tools

Tool	Description
loom_weave_block	Saves a raw reasoning block and its concise Warp summary. Frees cognitive load for the next step.
loom_view_tapestry	Returns all Warp summaries in chronological order for a thread.
loom_list_threads	Lists all known threads with block counts and timestamps.
loom_delete_thread	Deletes all blocks for a specific thread (requires confirmation flag).
loom_export_memento_dataset	Finalizes the session and exports all blocks to a versioned .jsonl file in Memento format.
loom_read_export	Reads an exported .jsonl file via MCP for in-protocol verification/retrieval.
loom_prune_check	Calculates reasoning density and advises when to call loom_weave_block.

Resources

URI	Description
loom://threads/{thread_id}/stats	Dynamic resource showing token savings for a thread.

---

Installation

Prerequisites

• Node.js 18+
• npm 9+

Setup

git clone https://github.com/Master0fFate/Loom-MCP.git
cd Loom-MCP
npm install
npm run build

The server uses SQLite and stores its database at ~/.loom-mcp/loom.db by default. Override this by setting the LOOM_DATA_DIR environment variable.

Token counts are computed using js-tiktoken (default model: gpt-4o-mini). Override model selection with LOOM_TOKENIZER_MODEL.

---

Configuration

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "loom-mcp": {
      "command": "node",
      "args": ["/absolute/path/to/Loom-MCP/dist/index.js"]
    }
  }
}

See also the included claude_desktop_config.json for a ready-to-use example.

---

Usage Workflow

1. Start a new reasoning task → pick a thread_id (e.g., "task-2024-001")
2. Reason deeply → call loom_weave_block with your raw analysis + concise summary
3. Before the next step → call loom_view_tapestry to review prior summaries
4. Periodically → call loom_prune_check to gauge context pressure
5. When finished → call loom_export_memento_dataset to create the SFT record

Warp Summary Quality Guide

High-density Warp summaries preserve decisions, assumptions, and outcomes:

• Good Warp: "Selected Dijkstra over BFS because weighted edges were introduced in step 3; complexity remains acceptable (O((V + E) log V)); next step is edge-case validation for disconnected nodes."

Low-density summaries lose reasoning value:

• Bad Warp: "Thought about graph options and picked one. Will continue."

---

Data Format

Exported .jsonl files (stored in ~/.loom-mcp/exports/) follow this schema:

{
  "id": "<uuid>",
  "thread_id": "<thread_id>",
  "exported_at": "<ISO timestamp>",
  "block_count": 3,
  "reasoning": "<think>\n<|block_start|>...<|block_end|>\n<|summary_start|>...<|summary_end|>\n</think>",
  "output": "<final answer delivered to the user>"
}

---