← Frameworks

Letta / MemGPTLetta / MemGPT

Letta, Inc.

开源Open source Apache2.0Apache2.0 有状态agentStateful agent 记忆分层Memory layers MemGPTMemGPT 服务器形态Server form partial lifecycle: active 核验Checked 2026-10-30

Letta（原 MemGPT）把 agent 做成持久实体（Apache-2.0）：agent 服务器常驻、记忆分层（core 自编辑 + archival 检索）、会话可中断续接。形态是它与其它框架最大的差异—— 它是服务器，不是库。

Letta (formerly MemGPT) makes agents persistent entities (Apache-2.0): a resident agent server, layered memory (self-edited core + retrieval-backed archival), sessions that interrupt and resume. Form is its biggest difference from other frameworks — it is a server, not a library.

适合与不适合When to use it, when not to

适合：长期记忆/「持久人格」agent、自托管 agent 服务器、MemGPT 设计路线落地。不适合：无状态一次性调用；只要库不要服务的架构。

Fits: long-memory 'persistent persona' agents, self-hosted agent servers, landing the MemGPT design. Not a fit: stateless one-shot calls; architectures wanting a library without a service.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-21

出品属地Origin: 国外出品foreign-made　Letta, Inc.（美国，MemGPT 论文团队公司化）。Letta, Inc. (US; the MemGPT paper team incorporated).
国内使用Direct access: 大陆可直连direct OK　letta.com 与 docs.letta.com 直连实测 200（2026-10-21）；GitHub 仓大陆直连波动（同日代理可达）。letta.com and docs.letta.com measured 200 direct (2026-10-21); the GitHub repo's mainland reachability fluctuates (200 via proxy the same day).
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-21 检索）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (searched 2026-10-21); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 各家模型经 provider 层接入（OpenAI/Anthropic/Gemini/本地等，以官方文档为准，未逐条清点）；它的核心叙事是有状态 agent：记忆分层（core/archival）由 agent 自己编辑管理（MemGPT 论文线）。
Various models connect via a provider layer (OpenAI/Anthropic/Gemini/local etc., per the official docs, not itemized one by one); its core narrative is the stateful agent: memory tiers (core/archival) edited and managed by the agent itself (the MemGPT paper line).
运行位置Where it runs: agent 服务器形态（区别于纯库）：起一个常驻服务（letta server），agent 与记忆在服务器里持久存在； SDK/CLI 连它。GitHub 25,290★，最近推送 2026-10-18（API 核验）。
An agent-server form (unlike a pure library): run a resident service (letta server); agents and memory persist inside the server; SDK/CLI connect to it. GitHub 25,290★, last push 2026-10-18 (API verified).
本地文件Local files: 自托管服务器把 agent 状态与记忆存在你的数据库（默认 SQLite，可配 Postgres）——数据在你侧； Letta Cloud 托管线是另一产品面。
The self-hosted server keeps agent state and memory in your database (SQLite by default, Postgres configurable) — data on your side; the Letta Cloud hosted line is a separate product surface.
关机后的任务Tasks after shutdown: 服务器形态自带常驻语义——这是它与其它框架大类的结构差异：agent 是持久实体，会话可中断续接（官方核心叙事）。
The server form carries residency semantics by itself — the structural difference from the other framework entries: the agent is a persistent entity, sessions can be interrupted and resumed (the official core narrative).
工具与扩展Tools & extensions: 工具/函数调用 + 文件读写类内置工具线；MCP 支持状态未逐条核验（官方文档有相关线，细节未核）。
Tool/function calling + built-in file-read/write tool lines; MCP support status not itemized (the official docs have related lines, details unchecked).
上下文与记忆Context & memory: 记忆管理就是它的产品本体：core memory（agent 自编辑）+ archival（外部检索存储）分层是 MemGPT 设计——跨会话记忆是它的默认面，不是附加组件。
Memory management is the product itself: the core memory (agent self-edited) + archival (external retrieval store) tiering is the MemGPT design — cross-session memory is its default surface, not an add-on.
权限与限制Permissions & limits: 服务器有鉴权配置面（细节未逐条核验）；工具执行（文件/命令类）发生在服务器环境—— 自托管机器的权限面要自己设防。
The server has an authentication configuration surface (details not itemized); tool execution (file/command kinds) happens in the server environment — the self-hosted machine's permission surface must be defended by you.
适合什么任务Best for: 适合：要「持久人格」的 agent（长期记忆+状态续接）、要自托管 agent 服务器、MemGPT 设计路线。不适合：无状态一次性调用（服务器形态是成本）；只要库不要服务（形态选择）。
Fits: agents needing a "persistent persona" (long-term memory + state resumption), self-hosted agent servers, the MemGPT design route. Does not fit: stateless one-shot calls (the server form is a cost); wanting only a library, not a service (a form-factor choice).

头号误解Top misconception

把 Letta 当「又一个 agent 库」——它是 agent 服务器形态（常驻+持久状态），选型时形态先对上Treating Letta as 'yet another agent library' — it is an agent-server form (resident + persistent state); match the form before choosing
改名史混淆：MemGPT（论文/旧项目名）→ Letta（公司化现名）——同一线，别当两个项目Confusion from the rename: MemGPT (paper/old project name) → Letta (the company's current name) — one line, not two projects
记忆在服务器数据库里——自托管备份与访问控制是运维责任，别按「框架不管数据」的直觉Memory lives in the server's database — self-hosted backups and access control are your operational duty; do not apply the 'frameworks don't manage data' intuition

未知项清单Known unknowns

MCP 支持状态（官方文档有相关线，细节未逐条核验）
服务器鉴权配置面细节（未逐条核验）
Letta Cloud 与自托管的能力差异表（未逐条对照）

MCP support status (the official docs have related lines; details not itemized)
Details of the server authentication configuration surface (not itemized)
A capability-difference table between Letta Cloud and self-hosted (not compared item by item)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-30，内容更新日content updated on 2026-10-21。

类型Type	名称Name	链接Link
repo	letta-ai/letta · 仓库（Apache-2.0；25,290★，最近推送 2026-10-18，GitHub API 核验）letta-ai/letta · repository (Apache-2.0; 25,290★, last push 2026-10-18, GitHub API verified)	https://github.com/letta-ai/letta
docs	Letta · 官方文档Letta · official docs	https://docs.letta.com
changelog	Letta · 官方 changelogLetta · official changelog	https://docs.letta.com/changelog
changelog	letta-ai/letta · Releases（changelog）letta-ai/letta · Releases (changelog)	https://github.com/letta-ai/letta/releases

实测记录Measured runs

本档未跑集成与性能实测；结论只覆盖 API 与官方文档披露的能力面。

This site has not completed measured runs. The test protocol is at protocols/framework/_protocol.md.