myelin-setup

name	myelin-setup
description	Set up Myelin knowledge graph memory for any project. Use when asked to "set up myelin", "index my code", "index my notes", "add memory to my agent", or "install myelin".

Myelin Setup

Purpose

Guide users through installing and configuring Myelin — a knowledge graph memory system for AI agents. After setup, agents get persistent memory via semantic search, auto-injected context, automatic session logging, and graph-based briefings. No manual boot or logging steps required — the extension handles it.

Required Configuration

This skill works without any pre-existing configuration. During setup, it creates the necessary infrastructure. If the user wants to index code repos or notes, they provide the paths interactively.

Instructions

Phase 1: Install & Initialize

1. Check if myelin CLI is already installed:

   myelin --version
   

2. If not installed, install from GitHub:

   npm install -g github:shsolomo/myelin
   

   Requires Node.js 24 (25+ is not yet supported). Native addons (better-sqlite3, sqlite-vec) compile during install — a C++ toolchain is required (Visual Studio Build Tools on Windows, Xcode on macOS, build-essential on Linux).

3. Set up the Copilot CLI extension (also initializes the graph database and downloads models):

   myelin setup-extension
   

   This bundles the extension, creates the graph DB at ~/.copilot/.working-memory/graph.db, installs runtime dependencies, and downloads the GLiNER NER model + embedding model (~660MB one-time download).

4. Restart Copilot CLI (or run /clear) to load the extension.

5. Verify the installation:

   myelin doctor
   

   Should show ✅ for graph database, schema, and extension. Follow any ⚠️ or ❌ recommendations.

Phase 2: Index Content

Ask the user what they want to index. There are three modes — any combination works:

Mode A: Code Repository

Index source code with tree-sitter. Supports C#, TypeScript, JavaScript, Python, Go, YAML, JSON, Bicep, PowerShell, and Dockerfile.

myelin parse /path/to/repo --namespace repo:my-project --embed

Namespaces partition the graph by source. Use a consistent naming convention: repo:name for code, docs:name for documents, vault:name for vaults. The --embed flag generates embeddings for semantic search.

Optional: specify a namespace to keep repos separate:

myelin parse /path/to/repo --namespace repo:my-project --embed

This creates File, Class, Method, Interface, Function nodes with defines and contains edges — the structural skeleton of the codebase.

To list indexed namespaces:

myelin namespaces

Mode B: Text Documents (Notes, Meeting Recaps, etc.)

Ingest text files with entity extraction and relationship edges:

myelin ingest /path/to/notes --namespace docs:notes --embed

Uses NER (GLiNER for high-precision zero-shot extraction, with regex/heuristic fallback if the model isn't available) to extract entities and create co-occurrence edges. GLiNER downloads automatically during myelin setup-extension.

Use --fast to skip embedding-based relationship classification (faster, proximity-only edges):

myelin ingest /path/to/notes --namespace docs:notes --fast

Mode C: IDEA Vault (Structured Notes)

For IDEA-method vaults (Initiatives, Domains, Expertise, Archive):

myelin vault /path/to/vault --namespace vault:main --embed

Mode D: Agent Memory (Structured Logs)

Agents log automatically via the extension's lifecycle hooks — no manual logging setup needed. The extension:

• Auto-detects the agent name on session start
• Auto-injects graph context before the first message
• Auto-logs a session summary on session end

For agents that want to log additional structured events during sessions, they can use the myelin_log tool or the CLI:

myelin agent log <agent-name> <type> "<summary>" --tag <topic>

Types: decision, action, finding, error, observation, handover

Phase 3: Consolidate & Embed

Run a full maintenance cycle to process agent logs and generate embeddings:

myelin sleep

This single command:

1. Discovers all agents with logs
2. Runs NREM consolidation (replay logs → extract entities → score salience → write to graph)
3. Runs REM refinement (global decay → pruning)
4. Generates embeddings for all nodes (required for semantic search)

For per-agent or per-phase control:

myelin consolidate --agent <agent-name>              # consolidate one agent
myelin consolidate --agent <agent-name> --phase nrem  # NREM only
myelin embed                                          # embeddings only

Recommend scheduling myelin sleep nightly for ongoing maintenance.

Phase 4: Verify

1. Run diagnostics:

   myelin doctor
   

2. Check graph contents:

   myelin stats
   

3. Test semantic search:

   myelin query "your search term here"
   

4. Test agent boot (what agents see at session start):

   myelin agent boot <agent-name>
   

Phase 5: Agent Integration

The extension handles most integration automatically:

• Auto-boot: Graph context is injected on session start (no manual myelin_boot call needed)
• Auto-log: Session summaries are logged on session end
• Per-message context: Relevant knowledge is surfaced with each user message

To customize an agent's interaction with myelin, add instructions to the agent's definition file:

myelin agent instructions <agent-name>

This generates logging instructions you can paste into the agent definition.

Constraints

• Do NOT delete log files. Consolidation reads logs but never removes them. Logs are the source of truth for graph rebuilds.
• Do NOT skip myelin sleep (or myelin embed) if the user wants semantic search — without embeddings, queries fall back to keyword-only FTS5 search.
• Native addon compilation requires a C++ toolchain (Visual Studio Build Tools on Windows, Xcode on macOS, build-essential on Linux).
• If myelin doctor reports issues, address those before troubleshooting other problems.

Examples

Full setup for a new project

# Install
npm install -g github:shsolomo/myelin

# Set up extension (auto-inits graph DB, downloads models)
myelin setup-extension

# Restart Copilot CLI, then:

# Index your codebase
myelin parse ./my-project --namespace repo:my-project --embed

# Ingest your notes
myelin ingest ./my-notes --namespace docs:notes --embed

# Consolidate + embed
myelin sleep

# Verify
myelin doctor
myelin query "cache performance"

Index multiple repos

myelin parse ./frontend --namespace repo:frontend --embed
myelin parse ./backend --namespace repo:backend --embed
myelin parse ./infra --namespace repo:infra --embed
myelin namespaces  # shows all three

Re-index after graph corruption

Logs are never deleted, so the graph can always be rebuilt:

rm ~/.copilot/.working-memory/graph.db
myelin init
myelin parse ./my-repo --namespace repo:my-repo --embed
myelin sleep