Letta Documentation — Building Stateful Agents with Persistent Memory
by Letta
The MemGPT lineage, productised: agents whose memory is a git-backed filesystem they edit themselves.
Overview
Letta is the successor to MemGPT, the UC Berkeley work on paging information in and out of an LLM context window, and positions itself as "the platform for building stateful agents" that "learn from experience and improve with use." Documentation is organised as Get started (Overview, Quickstart, Pricing, Self-hosting, Concepts), Platform (CLI, Desktop app, Cloud agents, ACP integration, Computers, App server), Configuration (Memory and dreaming, Skills, Mods, Permissions, Channels, Schedules, Secrets, Subagents, Models), Enterprise and Team, and Reference. The memory system is the substance. The current Agent SDK uses MemFS, "a git-backed memory filesystem" the agent can inspect and edit, which persists across conversations; /init bootstraps it by having the agent inspect your repository and ask about your working style, /remember teaches it explicitly, and /doctor audits placement, duplication and system-prompt token usage. "Dreaming" runs background subagents that review recent conversations, consolidate lessons and update memory without interrupting active work, optionally with a second background review pass before changes land. The earlier V1 SDK's labelled memory blocks (human, persona) and the portable Agent File (.af) export format are also documented. Agents reach users through Slack, Telegram, Discord, WhatsApp and Signal channels, a desktop app on macOS, Windows and Linux, chat.letta.com, or a TypeScript SDK. Pricing runs Free at $0/month with limited agents, Pro at $20/month for up to 20 stateful agents, an API plan at $20/month plus $0.10 per active agent per month and $0.00015 per second of tool execution, Teams Pro per seat, and custom Enterprise; you can bring your own model API keys on any tier. Note the repository has been restructured: the Apache-2.0 letta-ai/letta repo (~24.4k stars) is now a landing page with the V1 server preserved on an archive branch, and active development has moved to letta-ai/letta-code.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Documentation
- Cost
- Freemium
- Duration
- ~1 hour quickstart; ~5-7 hours across the full documentation
- Provider
- Letta
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Build agents whose state survives restarts instead of rebuilding context each session
- ✓Structure agent memory as a git-backed MemFS the agent can read and edit
- ✓Configure dreaming so background subagents consolidate lessons after each session
- ✓Audit memory growth and system-prompt token cost with the /doctor workflow
- ✓Export and re-create a complete agent using the Agent File (.af) format
- ✓Deploy agents to Slack, Discord, Telegram, WhatsApp and Signal channels
- ✓Decide between Letta Cloud, the local app server and full self-hosting
- ✓Delegate work to subagents and package reusable behaviour as skills
Highlights
- •Descends directly from the MemGPT paper, so the memory design has published research behind it
- •MemFS makes agent memory a git-tracked, human-inspectable filesystem rather than an opaque vector store
- •Dreaming is a genuinely uncommon feature — scheduled offline memory consolidation, not just retrieval
- •Apache-2.0 core with ~24.4k GitHub stars and a supported self-hosting path off the cloud
- •Documents its own migration honestly, including which SDK generation each feature belongs to
Who It’s For
Best For
- ✓Engineers building assistants that must remember users across weeks
- ✓Teams evaluating agent memory architectures beyond naive RAG over chat history
- ✓Developers who need a self-hostable stateful agent runtime
Prerequisites
- •Comfortable with Python or TypeScript and calling an LLM API
- •Understanding of context windows and why they constrain long-running agents
- •Basic command-line and git familiarity for the MemFS workflow
FAQ
What is Letta Documentation — Building Stateful Agents with Persistent Memory?
Documentation for the platform that grew out of the MemGPT research on giving LLM agents durable memory. Letta agents persist state across sessions, edit their own memory, and consolidate what they learned in background passes. The docs cover the CLI, desktop app, self-hosting and cloud, plus configuration of memory, skills, subagents, permissions and messaging channels for engineers building agents meant to run for months.
Is Letta Documentation — Building Stateful Agents with Persistent Memory free?
Letta Documentation — Building Stateful Agents with Persistent Memory offers free content, with paid options for certificates or premium features.
What level is Letta Documentation — Building Stateful Agents with Persistent Memory for?
Letta Documentation — Building Stateful Agents with Persistent Memory is aimed at a intermediate audience. Recommended background: Comfortable with Python or TypeScript and calling an LLM API, Understanding of context windows and why they constrain long-running agents, Basic command-line and git familiarity for the MemFS workflow.
How long does Letta Documentation — Building Stateful Agents with Persistent Memory take?
Expect roughly ~1 hour quickstart; ~5-7 hours across the full documentation. Most learners work through it at their own pace.
What will I learn from Letta Documentation — Building Stateful Agents with Persistent Memory?
You'll learn: Build agents whose state survives restarts instead of rebuilding context each session; Structure agent memory as a git-backed MemFS the agent can read and edit; Configure dreaming so background subagents consolidate lessons after each session; Audit memory growth and system-prompt token cost with the /doctor workflow; Export and re-create a complete agent using the Agent File (.af) format; Deploy agents to Slack, Discord, Telegram, WhatsApp and Signal channels; Decide between Letta Cloud, the local app server and full self-hosting; Delegate work to subagents and package reusable behaviour as skills.
Topics
Sources
This page was written from 4 sources, 2 on domains other than docs.letta.com.