Letta (formerly MemGPT)

Letta, formerly MemGPT, is a developer-focused platform for building stateful AI agents whose memory can persist beyond a single prompt, giving teams a practical way to create assistants that remember selected user, task, document, and workflow context across sessions while still exposing programmable controls through APIs, SDKs, hosted services, and self-hosted deployment patterns. Its core value is not simply longer chat history; it is an architecture for deciding which facts should live in active context, which should move to archival memory, and how an agent should retrieve or update that information during future work.

The platform is best understood as memory infrastructure for agent applications. Letta's public API documentation identifies a REST API with versioned v1 endpoints for agents, messages, tools, blocks, folders, files, archives, passages, models, MCP servers, runs, steps, feedback, conversations, and access tokens. That breadth matters because persistent memory becomes useful only when it is connected to the rest of the agent lifecycle: creating agents, attaching memory blocks, sending messages, running tools, searching archival memory, importing or exporting agents, and managing files or folders. The API reference also lists official client libraries for TypeScript and Python, with the TypeScript package shown as @letta-ai/letta-client and the Python package shown as letta-client, giving engineering teams two common implementation paths.

Several verifiable product facts define the current positioning. First, Letta supports a free entry point, with a Free tier listed at $0/month in the current record. Second, the current pricing record lists an API Plan at $20/month plus usage-based charges. Third, the same record lists $0.10 per active agent per month and $0.00015 per second for server-side tool execution, so production teams should model active-agent counts, tool runtime, and LLM usage rather than assuming a flat subscription covers all volume. Fourth, the API reference exposes explicit agent memory surfaces, including core-memory block endpoints and archival-memory passage endpoints, which validates the memory-first positioning. Fifth, the API docs list MCP server endpoints, including create, list, retrieve, update, delete, refresh, list tools, retrieve tool, and run tool operations, supporting the claim that Letta can participate in MCP-style tool workflows where configured. Additional current docs also show BYOK-style use with providers such as OpenAI, Anthropic, and OpenRouter for Letta Code, while Letta API examples demonstrate creating a stateful agent with memory blocks and an OpenAI model identifier.

Letta is strongest when an application needs continuity over time: support agents that remember unresolved issues, developer assistants that retain repository conventions, research agents that keep source notes, or personalized assistants that should reuse carefully selected preferences. It is less compelling for one-shot Q&A, short-lived workflows, or teams that primarily need general orchestration rather than durable state. Persistent memory also adds governance work. Teams need policies for what gets stored, how stale or wrong memories are corrected, how sensitive data is handled, and when a user or operator can delete retained context.

For buyers, the practical evaluation should focus on three questions. Can the team design reliable memory behavior for its domain? Does Letta's API and SDK model fit the product architecture? Do pricing and deployment options match expected agent count, tool execution volume, model usage, and compliance requirements? When those answers are positive, Letta offers a specialized memory layer that can make long-running agent products more coherent than stateless chat or retrieval-only systems.