Your agent finally remembers your dog’s name—and its own job

What it does

Letta is a platform for stateful LLM agents that remember context across sessions. It exposes a hosted agents API with Python and TypeScript SDKs, and bundles a separate CLI tool for local use. You configure agents with labeled memory blocks—such as a human profile and a persona—plus tool access like web search, and the system persists that state between messages.

The interesting bit

The project rebranded from MemGPT and now offers both a local CLI and a hosted API around the same stateful agent core. What stands out is the explicit memory model: you declare labeled blocks like human and persona upfront, turning agent identity and user context into structured state rather than buried prompt text.

Key highlights

• Agents rely on structured memory_blocks with user-defined labels to persist facts and identity
• Supports skills, subagents, and pre-built bundles for memory and continual learning
• Model-agnostic, though the maintainers rank Opus 4.5 and GPT-5.2 highest on their public leaderboard
• Available through a hosted API with Python and TypeScript SDKs, plus a separate local CLI project
• Over a hundred open-source contributors

Caveats

• The README repeatedly claims agents “learn and self-improve over time” but never describes the mechanism, leaving it unclear whether this means memory updates, automated skill refinement, or something else
• The relationship between this repository and the separate letta-code CLI repository is fuzzy; most local-runtime details and installation instructions live in the other project

Verdict

Adopt it if your agents need to remember user facts and prior actions across long sessions. Look elsewhere if you want a simple stateless chat wrapper or a single, self-contained repository instead of a split CLI/API codebase.