Letta

Overview

Letta is the production evolution of the MemGPT research project. Its core abstraction is the stateful agent: a long-lived entity with explicit memory tiers (core memory, archival memory, recall memory) and an inspectable state. The right pick when an agent has to remember things across sessions.

Best for

Best for
Anyone building persistent agents whose memory matters

Use cases

• Build a long-lived assistant that remembers user preferences over weeks
• Operate agents whose value compounds with accumulated context
• Replace ad-hoc 'history in prompts' with structured memory
• Inspect and edit agent memory directly

Notes

Why it matters

Most agent frameworks treat memory as ‘whatever fits in the context window.’ Letta makes memory a first-class system primitive. For long-lived agents, that distinction is the whole product.

How teams use it in production

Pin the structured memory schema to your domain. Audit core memory weekly. Use archival memory as the source of truth for facts the agent should know.

What to watch

Memory as a first-class agent primitive is becoming the next frontier. Letta, mem0, and Cognee are the live contenders.