claude-mem0-plugin
Persistent memory for Claude Code via mem0. Fully implicit — no manual commands.
Stores personal preferences, project architecture facts, and known issues
across projects using Qdrant (vector) and optional Neo4j (graph).
Install
# Add marketplace (one-time)
/plugin marketplace add wuffig-coding-solutions/agent-plugins
# Install plugin
/plugin install mem0@agent-plugins
# Set credentials (hooks)
export MEM0_API_KEY=your-key
export MEM0_USER_ID=your-username
Restart Claude Code. Done.
What it does
| Hook | When | Action |
|---|
SessionStart | Session begins | Load user prefs + project context |
UserPromptSubmit | Every prompt | Search mem0, inject relevant memories |
Stop | Every response | Extract + save new facts |
PostCompact | After /compact | Re-inject project context |
SubagentStart | Subagent spawned | Inject task-relevant context |
SubagentStop | Subagent done | Save subagent findings |
Notification | Claude waiting | Desktop notification |
Scopes
user_id: "your-username" → global preferences, cross-project patterns
agent_id: "project-name" → project stack, architecture, debug insights
agent_id is auto-detected from the current directory name. Override:
export MEM0_AGENT_ID=my-project
Self-hosted stack
cd scripts && docker compose up -d
Starts Qdrant on localhost:6333 and Neo4j on localhost:7687.
With Neo4j configured, mem0 automatically uses graph memory (Mem0ᵍ).
Configuration
| Variable | Default | Description |
|---|
MEM0_API_KEY | required | From app.mem0.ai or self-hosted |
MEM0_USER_ID | user | Your user identifier |
MEM0_AGENT_ID | auto (dir name) | Project identifier |
MEM0_TOP_K | 5 | Max memories per search |
MEM0_USER_THRESHOLD | 0.35 | Min score, user scope |
MEM0_AGENT_THRESHOLD | 0.25 | Min score, project scope |
MEM0_MAX_MESSAGES | 20 | Max messages saved per session |
NEO4J_URI | — | Enable graph memory |
NEO4J_USER | neo4j | Neo4j username |
NEO4J_PASSWORD | — | Neo4j password |
MCP server
The plugin connects to the mem0 gateway via Streamable HTTP with OAuth2.
Claude Code handles authentication automatically on first connect — no
tokens to manage. Claude can proactively call search_memory,
add_memories, and list_memories in addition to the automatic hooks.
Memory audit
Run /mem0:audit to review memory quality and get improvement suggestions
for the extraction prompt.
Docs
1270011
upstash
severity1