letta
Letta (formerly MemGPT) is the 'agent with its own memory' approach. Instead of wiring memory into a separate layer, letta IS the agent — it manages its own working memory, long-term memory, and reflection. If you want a single system that handles both the 'thinking' and the 'remembering' for long-lived assistants, letta is the top open-source choice.
- Long-lived CEO assistants that need to remember everything forever
- Customer support agents where forgetting a customer is unacceptable
- Teams that want an agent system, not a memory layer — less glue code
What you'll do
letta runs as a self-hosted server via Docker. You interact with it through its web UI, CLI, or Python SDK. Budget 15 minutes.
Before you start
- Docker installed and running
- An OpenAI or Anthropic API key
- Comfort with docker compose basics
Step-by-step install
- 011. Pull and run the server
One-line docker run. Letta exposes a web UI on port 8283 by default.
docker run -d \ --name letta \ -p 8283:8283 \ -v letta_data:/var/lib/letta \ -e OPENAI_API_KEY=sk-... \ letta/letta:latest
- 022. Open the web UI
Navigate to http://localhost:8283. You'll see the letta admin UI.
- 033. Create your first agent
Click 'New agent'. Give it a persona (e.g. 'senior chief of staff'), a human (your name and context), and pick a model. Save.
- 044. Chat with it — watch its memory grow
Start chatting. Letta's UI shows you the agent's core memory (short, always-loaded context) and its archival memory (long-term recall). As you talk, you'll see both grow.
Tip: This is the educational moment. Letta makes memory visible in a way no other tool does — you can literally watch what the agent decides to remember. - 055. Connect via MCP or SDK
For production: use the Python SDK to call letta from your own backend, or the MCP bridge (community plugin) to use letta from Claude Code or Cursor.
Your first 10 minutes
- 01Tell the agent about yourself, your company, your top 3 priorities this quarter.
- 02Close the chat. Reopen. Confirm it remembers.
- 03Watch the core memory section — it reshapes itself over the first 20 messages. Don't fight it; see what it decides.
- 04Give it a long-term task ('remind me in a week if I haven't touched fundraising').
- 05Add Cognition CLO on top for the retention layer across your team.
Troubleshooting
The Docker container exits immediately.
Check your OPENAI_API_KEY environment variable is set. Letta needs a valid model provider on startup.
The web UI is slow.
Give Docker more memory (Docker Desktop → Settings → Resources). Letta benefits from at least 4GB.
letta holds the knowledge. Cognition CLO models retention per employee per concept using a Weibull forgetting curve — so you see decay before it becomes a missed SOP or a failed audit.