Letta
Letta (formerly MemGPT) is an open-source framework for building and deploying stateful AI agents. Its defining feature is persistent memory management, supporting the writing of conversation history, project information, and other data into a structured knowledge base. A retrieval mechanism enables cross-session context sharing across multiple conversation turns, giving agents long-term memory capabilities. Letta exposes a server-side API and is suited for multi-agent collaboration, customer service, coding assistants, and other applications requiring persistent state management.
Core Facts
Asynchronous tool-calling mode allows the application layer to submit time-consuming operations for background execution while letting the model continue processing other subtasks that have no dependencies.
MemGPT has been renamed to Letta; it simulates the virtual memory mechanism of an operating system, allowing agents to manage information that exceeds the context window capacity within a limited context window.
Timeline (last 90 days)
Letta's core memory designs memory as structured state blocks with independent lifecycles, rather than natural language buried in the system prompt.
The target users of this Letta architecture are teams seriously wanting to embed AI into their R&D processes, not individual developers who just want autocomplete.
Stateful agent architectures bring operational costs, including memory update policies, contamination prevention and multi-agent concurrent write consistency.
The Letta tutorial emphasizes packaging agent configuration, memory structure and tool definitions into a portable form to support reuse locally, in CI and across team members.
The Letta tutorial mentions multi-agent shared state, i.e., multiple agents specializing in different responsibilities sharing the same memory foundation.
Letta allows capabilities such as reading files, running tests, checking repository status and calling internal APIs to be wrapped as custom tools for the agent to call.
Archival vector retrieval solves the memory capacity problem by vectorizing historical information into an archive and recalling it by semantic similarity.
The Letta tutorial positions core memory as frequently used working memory of limited size, and archival memory as a massive long-term archive retrieved on demand.
Core memory provides controllability: when memory is contaminated or outdated, it can be corrected directly without redesigning the entire prompt.
Core memory provides inspectability: developers can read out at runtime the key information the agent currently holds.
16 more timeline events
All Facts (20)
Asynchronous tool-calling mode allows the application layer to submit time-consuming operations for background execution while letting the model continue processing other subtasks that have no dependencies.
75%VerifiedMemGPT has been renamed to Letta; it simulates the virtual memory mechanism of an operating system, allowing agents to manage information that exceeds the context window capacity within a limited context window.
65%Unverified斯坦福大学的生成式智能体（Generative Agents）研究和Letta（原MemGPT）等项目也在探索跨会话持久记忆和系统级学习方向
85%UnverifiedMemGPT has been renamed Letta and proposed the concept of a hierarchical memory architecture, dividing memory into three levels: working memory, short-term memory, and long-term memory.
60%UnverifiedThe target users of this Letta architecture are teams seriously wanting to embed AI into their R&D processes, not individual developers who just want autocomplete.
50%UnverifiedLetta's core memory designs memory as structured state blocks with independent lifecycles, rather than natural language buried in the system prompt.
50%UnverifiedCore memory provides inspectability: developers can read out at runtime the key information the agent currently holds.
50%UnverifiedCore memory provides controllability: when memory is contaminated or outdated, it can be corrected directly without redesigning the entire prompt.
50%UnverifiedThe Letta tutorial positions core memory as frequently used working memory of limited size, and archival memory as a massive long-term archive retrieved on demand.
50%UnverifiedArchival vector retrieval solves the memory capacity problem by vectorizing historical information into an archive and recalling it by semantic similarity.
50%UnverifiedLetta allows capabilities such as reading files, running tests, checking repository status and calling internal APIs to be wrapped as custom tools for the agent to call.
50%UnverifiedThe Letta tutorial mentions multi-agent shared state, i.e., multiple agents specializing in different responsibilities sharing the same memory foundation.
50%UnverifiedThe Letta tutorial emphasizes packaging agent configuration, memory structure and tool definitions into a portable form to support reuse locally, in CI and across team members.
50%UnverifiedStateful agent architectures bring operational costs, including memory update policies, contamination prevention and multi-agent concurrent write consistency.
50%UnverifiedLetta (MemGPT) has 25,091 GitHub stars, offers a web app and CLI, supports OSS server self-hosting, starts at $0 then Pro at $20/month, and rewrites memory in the background via sleep-time agents.
50%UnverifiedThe observer pattern (an independent LLM traverses and reads conversations and proactively writes structured memories) has been adopted by at least seven products, including mem0 (single-pass extraction plus Dream consolidation), Letta (sleep-time agents), Zep (Observations on the Flex Plus tier) and mcp-memory-service (autonomous consolidation and insight cards).
50%UnverifiedSelf-hosted memory is almost always free (the OSS apps of mem0, Graphiti, Letta, cognee, Hindsight, mcp-memory-service and supermemory), or locked behind enterprise contracts (Zep BYOC, mem0 local deployment, cognee BYOC), and the only public self-serve tier that includes self-hosting is supermemory's Scale at $399/month.
50%UnverifiedLetta provides editable memory blocks and conversation records, plus a Git-versioned memory filesystem, all packed into a stateful Agent runtime; adopting it means adopting its entire runtime.
50%UnverifiedLetta provides editable memory blocks, a transcript, and a Git-based memory filesystem, all inside a stateful Agent runtime; what it replaces is the Agent runtime, not just the memory layer.
50%UnverifiedOMPS is a vendor-neutral portability data model project that provides memory-item.schema.json and deletion-marker.schema.json.
50%