Letta
Agents with stateful memory (formerly MemGPT)
- Best for
- Agents that need persistent, long-term memory
- Pricing
- Open-source server; Letta Cloud has usage-based pricing
- Stack layer
- AI Stack
Live preview of letta.com
Overview
Letta (formerly MemGPT) was built around one observation: most agents lose context the moment a session ends. Letta makes memory the first-class primitive — core memory, archival memory, conversation history — all persisted automatically.
This is what you want when the use case is "an AI that knows the user over months" — long-running assistants, customer support bots that remember prior tickets, internal knowledge agents that learn the company.
The hosted Letta Cloud handles persistence and scaling; the open-source server is self-hostable for security-conscious deployments.
Key Features
Core Memory
Always-on context the agent edits over time
Archival Memory
Vector-backed long-term storage for retrieval
Persistent Identity
Agents that survive process restarts
Tool Use
Function calling with the standard schemas
Model Agnostic
Run on OpenAI, Anthropic, or local models
Open Source
Apache 2.0 with a hosted cloud option
Why We Recommend Letta
Stateful memory is the missing piece in most agent stacks. Letta gives it to you without rolling your own vector + summary pipeline.