What is Letta?
Letta, formerly MemGPT, builds stateful agents: agents that keep memory, identity and conversation history across sessions instead of starting blank each time. The project has been rebuilt. The original Python API server in letta-ai/letta (about 25,300 stars) was moved to an archive branch on 16 August 2026 and is no longer supported; active development is in letta-ai/letta-code, a TypeScript harness with a terminal interface, an App Server for self-hosted agents, and channels for Slack, Telegram and Discord.
Agents store their context in MemFS, a git-backed memory filesystem they read and rewrite. Background "dreaming" subagents review recent conversations and update memory between tasks, and agents can call subagents, run on schedules and work across several machines. You bring your own model keys (OpenAI, Anthropic and others) or use Letta Cloud's models.
Install is npm install -g @letta-ai/letta-code, or uv and pipx. The npm package recorded about 483,000 downloads in the 30 days to 29 October 2026, and v0.34.38 shipped on 2 November 2026.
Is Letta worth it?
Letta is now a personal and team agent harness rather than the memory server many developers adopted. The MemGPT paper's idea, an agent that manages its own context, survives in a more literal form: memory lives as files in a git repository the agent edits, so every change is versioned and can be synced to your own GitHub remote. That is easier to audit than the opaque vector stores most memory layers use, and it suits agents that run for weeks on one codebase or inbox.
The rebuild is the main risk. The Python server that applications called over an API was archived on 16 August 2026, receives no fixes or security updates, and the README warns against using it in production. AgentFile (.af) export and import, once the project's portability format, have been removed from Letta Code. Teams with a service on the old server face a migration to the App Server or the TypeScript Letta Agent SDK, and the stars on the new repository, about 3,800, show how much of the community has moved so far.
The pace cuts both ways. GitHub lists 100 releases between 8 August and 2 November 2026, close to one a day, which means fast fixes and a moving interface. Open issues include memory isolated per agent so new agents start blank, and a cloud runtime that drops reasoning output on many turns.
Costs depend on where state lives. Run locally with your own keys and Letta costs nothing beyond tokens. Letta Cloud is the default on first launch; its pricing page estimates casual coders at about $100 a month and power users at $200 or more, well above Pro's $20. The repository's privacy policy says hosted-service requests and responses may be used to improve services, including training future models, so teams with sensitive data should run the local backend. Mem0 and Graphiti are narrower memory layers that slot into an existing agent; Letta asks you to adopt its whole harness.
Where it earns its place
A solo developer maintaining several client codebases
Keep an assistant that remembers each project's conventions between sessions
Run letta in each repository, use /init to build project memory and /remember for rules such as package manager choice; memory commits to MemFS and can sync to a private GitHub repo. With a ChatGPT or Z.AI coding plan connected, the harness itself costs nothing.
A three-person operations team at a small agency
Have one agent answer routine questions in Slack and Telegram with shared context
Teams Pro at $20 a seat is $60 a month for shared agents and access controls; channels connect the same agent to Slack and Telegram, and scheduled heartbeats let it post summaries without being asked.
A backend team with a product built on the Letta V1 server
Decide whether to migrate or move to another memory layer
The archived server gets no security fixes, so staying put is not an option for production. Exporting memory blocks and archival data first keeps the choice open between Letta's App Server and the Agent SDK, Mem0 or Cognee.
Before you commit
- Setup
- npm install -g @letta-ai/letta-code (or uv tool install letta, Python 3.9+), run letta in a project directory, choose local or Letta Cloud, and connect model keys with /connect.
- Integrations
- OpenAI, Anthropic and other providers through your own keys; ChatGPT, Kimi and Z.AI coding plans; Slack, Telegram, Discord and custom channels; skills from GitHub, ClawHub and Hermes; a GitHub Actions runner; the TypeScript Letta Agent SDK.
- Data and privacy
- Apache 2.0. The local backend keeps state under ~/.letta on your machine. Letta Cloud stores memory, identity and conversations; the repository privacy policy says hosted requests and responses may be used to improve services, including training models. No certifications are published in the sources we read.
- Cost at scale
- Free locally apart from tokens. On Letta Cloud, Letta's own guide puts casual coding at about $100 a month and heavy use at $200 or more; the API plan adds $0.10 per active agent a month and $0.00015 per second of tool execution.
- Limits
- Cloud memory repositories are capped at 5 MiB on Free and 50 MiB on Pro; remote computers and secrets need a Letta Cloud sign-in; automatic dreaming is off by default on native Windows.
What’s good
- Memory stored as git-versioned files the agent edits
- Runs locally with your own model keys at no licence cost
- Background memory consolidation between tasks
- Same agent reachable from CLI, browser, Slack, Telegram and Discord
- Apache 2.0 licence and very frequent releases
What’s not
- Original Python server archived and unsupported
- AgentFile export and import removed
- Letta Cloud is the default backend at first launch
- Hosted-service data may be used for training under the privacy policy
- Near-daily releases mean a moving interface
Best for
Developers who want long-running personal or team agents with inspectable, versioned memory and are happy to adopt a whole harness.
Letta alternatives
Letta: common questions
- Is Letta worth it?
- Letta still has the most considered ideas about agent memory, but rebuilding the product mid-flight left its earlier users to migrate and newcomers with a fast-moving target. We rate it 3.7 out of 5.
- How much does Letta cost?
- Apache 2.0, free to run locally with your own model keys. Letta Cloud: Free (3 agents, 5 MiB per memory repository); Pro $20/month (20 agents, 50 MiB, model quota); API plan $20/month plus $0.10 per active agent a month and $0.00015 per second of tool execution, tokens extra; Teams Pro $20/seat/month; Enterprise custom.
- What are the downsides of Letta?
- Original Python server archived and unsupported; agentFile export and import removed; letta Cloud is the default backend at first launch; hosted-service data may be used for training under the privacy policy; near-daily releases mean a moving interface.
- Who is Letta best for?
- Developers who want long-running personal or team agents with inspectable, versioned memory and are happy to adopt a whole harness.
- How much does Letta cost for a team or in production?
- Free locally apart from tokens. On Letta Cloud, Letta's own guide puts casual coding at about $100 a month and heavy use at $200 or more; the API plan adds $0.10 per active agent a month and $0.00015 per second of tool execution.
- Is Letta safe for company data?
- Apache 2.0. The local backend keeps state under ~/.letta on your machine. Letta Cloud stores memory, identity and conversations; the repository privacy policy says hosted requests and responses may be used to improve services, including training models. No certifications are published in the sources we read.
- What are the best Letta alternatives?
- The closest alternatives we have reviewed are Mem0, Graphiti, Cognee and LangGraph.
Answers checked Nov 2, 2026.