mem0

Why It Exists

AI agents are stateless by default. Every time a context window resets, everything you established — your conventions, decisions, preferences — is gone. Mem0 fixes that with a memory layer that persists across sessions, agents, and team members.

Mem0 sits between your agent and its context window. On every interaction, it automatically extracts facts worth keeping (preferences, decisions, patterns you've established) and stores them. On subsequent sessions, it injects the relevant memories back — so your agent "remembers" your coding style, your project conventions, and what you already decided last week.

Setup — Option A: Cloud (~2 min)

Sign up at app.mem0.ai (opens in a new tab), grab your API key, then:

Setup — Option B: Self-Hosted with Docker (~10 min)

1git clone https://github.com/mem0ai/mem0
2cd mem0/server
3docker compose up -d
4# Default endpoint: http://localhost:8080

Setup — Option C: MCP Integration (recommended for Claude Code & Cursor)

Add to your claude_desktop_config.json or mcp_settings.json:

1{
2  "mcpServers": {
3    "mem0": {
4      "command": "npx",
5      "args": ["-y", "@mem0/mcp-server"],
6      "env": {
7        "MEM0_API_KEY": "your_api_key_here"
8      }
9    }
10  }
11}

For local deployment, replace the env with MEM0_BASE_URL: http://localhost:8080.

What the Skill Does

• Auto-retrieval on session start— searches for relevant memories and injects them before proceeding
• Auto-storage during work— when a decision, preference, or pattern is established, the agent stores it without prompting
• Smart what-to-store rules— captures project decisions, naming conventions, and preferences; ignores temporary debugging steps and obvious facts
• Conflict resolution— surfaces conflicts between stored memories and what you say, asks which is current

User-Initiated Commands

• "Show me what you remember" — broad search, lists relevant memories
• "That memory is outdated, update it" — deletes the old, adds the corrected version
• "Forget [X]" — searches for X, confirms before deleting
• "Remember that [preference]" — stores immediately with context tags
• "What do you know about [topic]?" — targeted search and summary

Best Practices

• Be specific in storage — "Uses Tailwind v4 with @theme tokens" beats "uses Tailwind"
• Store the WHY — "Decided to migrate to Bun (April 2026) for speed gains"
• Tag with project context — when working across repos, prefix memories with the project name
• Review periodically — ask the agent to show all memories every few months and prune stale ones

Skill File

Drop the MEM0.md skill into your project to encode this behavior — copy or download from the Skills tab.

Troubleshooting

• Agent not retrieving memories automatically?Ensure the MCP server is listed in active connections. In Claude Code: /mcp
• Memories seem irrelevant?Improve search terms — use project name + specific domain instead of generic queries
• Too many memories cluttering context?Run a cleanup: "Show me all memories, I'll tell you which to remove"

Source

Source on GitHub (opens in a new tab) · Documentation (opens in a new tab)