Letta

Sweet spot: a product where the value proposition is "it remembers you." Coaching apps, companion agents, long-lived customer-service identities, research assistants. Letta solves the hardest part — memory that actually holds up — and does it as an opinionated, well-documented framework rather than asking you to invent it. Failure modes. If your agent is short-lived or your memory needs are really just "retrieve from a document store on each query," Letta is overhead you do not need — a RAG pipeline will be simpler and cheaper. The MemGPT approach is powerful but opinionated; if your memory shape is graph-natural or episodic, you may fight the abstraction. And memory quality is not just the framework's job — weak models produce unreliable memories regardless of scaffolding. What to pilot. Build one agent with a memory-centric value prop. Run it for two weeks on your own usage. Week 2, test it on things you mentioned in Week 1 — preferences, goals, specific facts. If recall is accurate and feels useful, Letta has earned its place; if the memory is noisy or irrelevant, investigate whether the issue is framework, model, or your prompts before committing.

Letta (formerly MemGPT) is an open-source framework for building agents that have persistent memory across sessions. The core idea, from the MemGPT paper: treat LLM context as an operating-system memory hierarchy — a fast "main context" the model sees directly and a larger "external context" the agent can page in and out via tool calls. Combined with an agent loop that proactively writes summaries and retrieves on demand, this gives agents something much closer to durable memory than a typical conversation buffer. Letta exposes this as a server (Docker or cloud) with a REST / Python SDK. Agents are first-class entities: they persist, they can be shared across users, and they accumulate memory over time. The server handles memory consolidation, archival, and retrieval; you just call the agent and it remembers. The company also offers Letta Cloud — a hosted version with persistent storage, API-key access, and team features — and an Agents Development Environment (ADE) web UI for inspecting agent memory and tracing steps. Apache-2-licensed, based on peer-reviewed research, and actively maintained by the original MemGPT authors. It is the cleanest answer in open source to "how do I build an agent that remembers things a month later."