Letta AI

Overview

Letta AI (formerly MemGPT) is a company and open-source framework specializing in persistent agent memory architecture. Their core contribution is the MemGPT design: a tiered memory system that gives LLM agents effectively unlimited memory by managing in-context memory and external storage automatically, analogous to how an operating system manages RAM and disk.

Letta AI(前身为 MemGPT)是一家专注于持久化智能体记忆架构的公司及开源框架。其核心贡献在于 MemGPT 设计:一种分层记忆系统,通过自动管理上下文记忆与外部存储,赋予 LLM 智能体有效无限的记忆能力,其运作机制类似于操作系统对内存(RAM)和磁盘的管理。

Key Facts

  • Formerly: MemGPT (academic project from UC Berkeley)
  • Became: Letta AI (commercialized entity)
  • Core Product: Persistent agent memory system
  • Open Source: Yes (active open-source project alongside commercial offering)
  • Key Innovation: OS-inspired memory paging for LLM agents
- **原名**:MemGPT(加州大学伯克利分校的学术项目)- **现名**:Letta AI(商业化实体)- **核心产品**:智能体持久化记忆系统 - **开源**:是(与商业产品并行的活跃开源项目)- **核心创新**:受操作系统启发的 LLM 智能体内存分页技术

Technical Architecture

Letta implements a tiered memory model:

  1. Core memory (in-context): high-priority, always-available information about user and agent persona
  2. Archival memory (external vector store): searchable long-term memory, retrieved on demand
  3. Recall memory (conversation history): searchable conversation archive
Letta 实现了一种分层记忆模型:1. **核心记忆**(上下文内):关于用户和智能体人设的高优先级、始终可用的信息 2. **归档记忆**(外部向量存储):可搜索的长期记忆,按需检索 3. **回溯记忆**(对话历史):可搜索的对话归档

The agent manages its own memory through explicit memory operations: core_memory_append, archival_memory_insert, archival_memory_search. This makes memory management an explicit, auditable part of agent behavior rather than an implicit side effect.

Agent 通过显式内存操作管理自身的内存:`core_memory_append`、`archival_memory_insert` 和 `archival_memory_search`。这使得内存管理成为 Agent 行为中显式且可审计的一部分,而非隐式的副作用。

Significance

Letta’s approach solves the “context window amnesia” problem that makes standard agents unreliable for long-running tasks. By making memory explicit and persistent, agents can sustain coherent behavior across sessions spanning days or weeks — a prerequisite for Deep Agent use cases.

Letta 的方法解决了“上下文窗口遗忘”问题,该问题导致标准智能体在长期运行的任务中不可靠。通过将记忆显式化和持久化,智能体能够在跨越数天或数周的会话中保持连贯的行为——这是深度智能体(Deep Agent)应用场景的先决条件。