mem0

What you'll do

mem0 is the production default for giving your AI agents long-term, multi-user memory. Most teams reach for it once they outgrow a single-user setup. It's mature, well-documented, and has both Python and TypeScript SDKs. Budget 20 minutes.

Before you start

• Python 3.10+ (or Node.js 20+ if you prefer the TS SDK)
• An OpenAI API key (or compatible — mem0 uses an LLM for extraction)
• Optional: a vector database (mem0 ships with a local one that works fine for <100 users)

Step-by-step install

1. 01
   1. Install the SDK

   Pick Python or TypeScript based on whatever backend you're already running.

   # Python
   pip install mem0ai
   
   # OR TypeScript
   npm install mem0ai

2. 02
   2. Set your API key

   mem0 uses OpenAI (or any OpenAI-compatible model) to extract memories from conversations. Export your key as an environment variable.

   export OPENAI_API_KEY=sk-...

3. 03
   3. Try it in a script

   Start with a 10-line Python script to verify mem0 is working end-to-end.

   from mem0 import Memory
   
   m = Memory()
   
   # Add a memory for a specific user
   m.add("I prefer dark mode and pair programming", user_id="amara")
   
   # Retrieve it later
   results = m.search(query="amara's preferences", user_id="amara")
   print(results)

   Tip: Every call is scoped by user_id — that's the unit of memory isolation. One scope per end-user, employee, or tenant.
4. 04
   4. Wire it into your agent loop

   Wherever your agent receives a user message, call m.search() to pull relevant context and prepend it to the prompt. Wherever it generates a response or the user commits a decision, call m.add() to store it. See the mem0 docs for framework integrations (LangChain, LlamaIndex, raw OpenAI).

5. 05
   5. Or: use the MCP server for Claude Code / Cursor

   If you want mem0 behind an MCP-speaking agent instead of your own backend, install the MCP bridge and point Claude at it.

   {
     "mcpServers": {
       "mem0": {
         "command": "npx",
         "args": ["-y", "@mem0/mcp"],
         "env": {
           "OPENAI_API_KEY": "sk-...",
           "MEM0_USER_ID": "your-org-or-user-id"
         }
       }
     }
   }

6. 06
   6. Upgrade to the hosted or self-hosted server when you scale

   For production, point mem0 at a persistent Postgres or a managed vector DB (Pinecone, Weaviate, Qdrant). The mem0 docs walk through each backend. For sub-100-user teams, the default local store is fine.

Your first 10 minutes

1. 01Decide your scoping strategy: one user_id per employee? per end-customer? per tenant? This is the single most important design decision.
2. 02Wire mem0 into one real workflow end-to-end (e.g. your CX agent) before expanding. Prove retrieval quality on one flow.
3. 03Add retrieval logging — you want to see which memories are being pulled and whether they're relevant.
4. 04Connect Cognition CLO — mem0 stores, CLO models which concepts are decaying per user.
5. 05Set up a weekly review of the memory graph for a sample user. mem0's LLM extraction sometimes drops things you want kept.

Troubleshooting