Letta Agent is a personalized AI agent for growing your own “digital coworker.” It traces its roots to MemGPT, a research project born at UC Berkeley’s Sky Computing Lab, and it accumulates conversations and work records in a memory layer called MemFS. Unlike typical chat AI, where context resets with every session, Letta Agent learns how you work and what you prefer the more you use it ─ and the memory carries over even when you switch models from OpenAI to Anthropic. Its core CLI, Letta Code, is published as open source under Apache-2.0, and you can reach the same agent through desktop apps for macOS / Windows / Linux, the CLI, an SDK, or the browser.
Key Features
- Persistent memory through MemFS: Conversations, skills, system prompts and other agent internals are kept in a Git-tracked form. You can sync them with a GitHub repository and review how the memory changed as a diff. Switching models does not wipe the memory
- Continual learning as you use it: The agent rewrites its own context and skills during idle periods (sleep-time compute) to improve accuracy. No model retraining is involved ─ only your own agent gets tuned to you
- Build a persona with custom skills: Skills can be defined at three scopes ─ global, per project, and per agent ─ and you can write your own or install ones published elsewhere. You can also modify the harness itself to change behavior
- Letta Code, an open-source CLI: Installed with
npm install -g @letta-ai/letta-codeand licensed Apache-2.0. Register providers such as OpenAI and Anthropic with/connect, then switch models with/model - Direct it from the channels you already use: It connects to Slack, Discord, Telegram and the browser (chat.letta.com works on mobile), so you can hand off work without opening a terminal. Scheduled runs and work on remote devices are supported as well
- Multi-agent setups: An agent can call other agents as subagents, letting you split responsibilities across a team of agents
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Free | $0 | Up to 3 agents, limited Letta Auto usage, bring your own API keys |
| Pro | $20 / month | Weekly and monthly Letta Auto quota, pay-as-you-go overages, up to 20 stateful agents |
| API | $20 / month + usage | Unlimited agents. $0.10 per active agent per month, $0.00015 per second of tool execution, API key authentication |
| Teams Pro | $20 / seat / month | Team collaboration, shared agents with access control, Letta Auto quota |
| Enterprise | Contact sales | Volume-based pricing, increased quotas, role-based access control, SAML / OIDC SSO, dedicated support |
Every plan lets you bring your own API keys. Note that LLM usage is billed separately depending on the model you choose.
Pricing is current as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- Memory persists, so you do not have to restate your context every time; the agent gradually becomes a counterpart that understands your situation
- MemFS is Git-based, so you can inspect what the agent remembered and rewrote as a diff ─ it is less of a black box
- The core CLI is open source under Apache-2.0, so you can modify the harness yourself
- You can switch models without throwing away the memory, which makes it easier to avoid lock-in to a single vendor
- Many entry points ─ Slack, Discord, Telegram, browser, desktop app ─ make it easy to fit into your daily routine
- The free plan is usable for evaluation if you bring your own API keys
⚠️ Cons
- Memory, skills and multi-agent setups add up to a lot of concepts; moving over from a plain chat AI takes some learning
- The free plan caps you at 3 agents, so serious use assumes moving to a paid plan
- LLM usage is billed on top of the subscription, which makes the total cost harder to predict
- Released in April 2026, it is a relatively new service and its features and specifications are changing quickly
- How to manage and prune memory as it accumulates over time is left to the user’s own operational design
Comparison with Similar Services
| Item | Letta Agent | Claude Code | OpenAI Codex CLI | Cursor |
|---|---|---|---|---|
| Provider | Letta | Anthropic | OpenAI | Anysphere |
| Focus | General agent with memory that grows | Delegating coding from the terminal | Delegating coding from the terminal | AI-integrated editor |
| Persistent memory | Built in via MemFS | Supplemented by project/memory features | Limited | Limited |
| Model choice | Multiple providers, switchable | Anthropic models | OpenAI models | Multiple models selectable |
| License | Apache-2.0 (Letta Code) | Proprietary | Open source | Proprietary |
| Entry points | CLI, app, Slack and others, browser | CLI, IDE extensions | CLI, IDE extensions | Editor |
Who Is It For
- People tired of re-explaining the same background every time, who want a counterpart that remembers their working context
- People who want to compare and move between models without being tied to one AI vendor
- Developers who want to shape agent behavior themselves and do not mind working with open source
- Teams that want to hand off work from everyday channels such as Slack and Discord
- People who want AI in a continuing role, including scheduled runs and work on remote devices
Summary
Letta Agent aims to shift how you work with AI ─ from “a chat that resets every time” to “a coworker that grows as you use it.” Persistent memory via MemFS and an open-source CLI form the core, and a design where memory survives a model switch is a practical strength while AI tooling keeps churning. There are many concepts to absorb, so a good starting point is the free plan with your own API keys, raising a single agent on your daily work.