advanced-reasoning-mcp: Instructions file for Codex

AGENTS.md — Advanced Reasoning MCP

Operational guide for AI agents using the advanced-reasoning MCP server. This file tells you when and how to call each tool. For human/dev setup, see README.md.

What this server gives you

A persistent reasoning workspace on top of sequential thinking:

• A reasoning loop (advanced_reasoning) with confidence, reasoning-quality, meta-cognition, hypotheses, evidence, revisions, and branches.
• A graph memory of your thoughts, organized into named libraries, searchable by TF-IDF relevance.
• SystemJSON: structured, searchable documents for reusable workflows, instructions, and domain knowledge.

Everything persists to memory_data/ on disk and is reloaded on restart, so memory survives across sessions and conversations.

Mental model

library  ──contains──>  sessions + memory nodes (graph)
session  ──groups──>    a connected chain of thoughts toward one goal
node     ──links to──>  other nodes via builds_on / challenges edges
SystemJSON ──separate──> document store for workflows & reference data

A node ID looks like node_1753683441451_a1b2c3d4. A session ID looks like session_<ts>_<rand>. You wire the reasoning graph by passing prior node IDs into builds_on / challenges on later thoughts.

The core workflow

1. Pick or create a library for the topic (optional but recommended for multi-topic work). Default library is cognitive_memory.
   • create_memory_library { library_name } then switch_memory_library, or
   • just keep working in the current one.
2. Create a session for the goal: create_reasoning_session { goal }. Keep the returned session_id.
3. Reason step by step with advanced_reasoning, passing session_id every call so thoughts persist. Capture the returned nodeId of each step.
4. Link thoughts: on later steps, put earlier nodeIds in builds_on (this step extends them) or challenges (this step contradicts them).
5. Recall earlier insight anytime with query_reasoning_memory { query }.
6. Save reusable procedures/data as SystemJSON for future sessions.