memgraph

Skill Discovery new in v1.1

Before starting a task, agents run memgraph recommend "<task>" and get ranked skill recommendations. The scoring algorithm uses:

• • IDF weighting — rare terms score higher than common terms
• • Accent folding — French/Spanish/German queries work (e.g. "déclaration impôts" → ja-ma-impots)
• • Project name boost — "add JWT to tau" boosts tau-maintenance
• • Extended content matching — first 1000 chars of SKILL.md indexed
• • Graph boost — skills connected to high-scoring skills get +5%
• • Stop words + stemming — "tests" matches "test", "monitoring" matches "monitor"

# Recommend skills for a task
memgraph recommend "deploy traefik with SSL" --json

# Search by keyword
memgraph query "proxmox backup" --json

# Get related skills
memgraph related "jar-rbm21-manage" --json

One-Command Onboarding new in v1.1

Run memgraph setup in any repo and it:

• • Auto-discovers all 4 skill directories (~/.agents/skills, ~/.claude/skills, ~/.config/devin/skills, ~/.codeium/windsurf/skills)
• • Ingests 212 skills into the graph
• • Upserts AGENTS.md with memgraph usage instructions
• • Creates .devin/skills/memgraph-usage/SKILL.md

memgraph setup
# → Ingested skills from 4 directories into graph
# → Updated AGENTS.md with memgraph usage instructions
# → Created .devin/skills/memgraph-usage/SKILL.md

MCP Server new in v1.7

Agents like Claude Code auto-discover memgraph's capabilities via the Model Context Protocol — no CLAUDE.md rules needed. memgraph mcp runs a stdio JSON-RPC 2.0 server. Two-tier tool design: 8 core tools (always advertised) + 1 admin meta-tool for 15 rare operations. Total overhead: ~1158 tokens per API call.

# Register with Claude Code
claude mcp add memgraph -s user -- ~/.local/bin/memgraph mcp

# Core tools agents see:
# memgraph_projects, memgraph_recall, memgraph_read,
# memgraph_save, memgraph_list, memgraph_edit,
# memgraph_delete, memgraph_recommend

# Admin meta-tool (15 rare ops):
# memgraph_admin {command: "status|profile|query|...", args: {...}}

Project Registry new in v1.6

Stable project identification via git remote URLs. A registry (~/.memgraph/projects.json) maps human-readable names to memory paths. Work from any directory with --project <name>. Orphaned scopes can be rebound with memgraph attach.

# Register current repo under a name
memgraph attach myproject

# Access from any directory
memgraph recall "deploy" --project myproject

# List all projects
memgraph projects

Sectioned Memories new in v1.5

Memories support [slug] section markers. Recall returns a compact section index (slug + preview + line range) — agents fetch only the section they need with memgraph read <id> <slug>. Reduces token consumption by 60-90% for large memories.

# Save a sectioned memory
memgraph remember "[deploy] Use systemd timers [backup] rsync to /backup"

# Recall returns section index, not full content
memgraph recall "deploy"
# → Memory 123 — "ops" (score: 5.2)
#   [deploy]  L2-3  "Use systemd timers"
#   [backup]  L4-5  "rsync to /backup"

# Read only the section you need
memgraph read 123 deploy

Galaxy Visualization

Project clusters spiral out from a galactic core. Namespace nodes are golden core stars; skills are colored stars sized by degree. 1500 background stars, nebula dust, additive blending — your skills as a Milky Way.

• • Drag stars to reposition overlapping nodes
• • Click a star to read its full content
• • Search + Enter to fit visible nodes
• • Copy Visible Paths to clipboard
• • Toggle references / similar / shared-keyword edges

HTTP API

All graph commands available via HTTP, returning the same JSON format as the CLI.

# Search the graph (same scoring as CLI)
curl "http://localhost:8080/api/search?q=traefik"

# Get a specific skill as JSON
curl "http://localhost:8080/api/nodes/traefik-cloudflare-setup"

# Get the full graph
curl "http://localhost:8080/api/graph"

# Trigger a sync
curl -X POST "http://localhost:8080/api/sync"

Agent Bridges

Generate integration files for Claude Code, OpenCode, and Copilot. Each bridge includes skill discovery commands and memory commands.

memgraph bridge claude-code    # → .claude/CLAUDE.md
memgraph bridge opencode        # → .opencode/memory.json
memgraph bridge copilot         # → .copilot/settings.json

Graph Construction

Ingests both SKILL.md files and plain .md files. Builds edges from three relation types:

• • references — skill mentions another skill by name
• • similar — TF-IDF pseudo-semantic similarity
• • shared-keyword — overlapping tags/keywords

Quick Start

# Download the binary
chmod +x memgraph && sudo mv memgraph /usr/local/bin/

# One-command onboarding (auto-discovers all skill dirs)
memgraph setup

# Find skills for a task
memgraph recommend "deploy traefik with SSL" --json

# Launch the galaxy UI
memgraph serve --port 8080
# Open http://localhost:8080