Most agent CLIs are amnesiacs: every session starts from zero, and the only thing that survives is whatever you paste back in. The wager here is the opposite — an agent gets useful precisely when it is allowed to keep things. Memory, skills, the system prompt, and even parts of the harness are treated as persistent state, written to disk and tracked in git like source code, so the agent you talk to next week is shaped by what it did last week.
What Sets It Apart
- Memory as a filesystem (MemFS): an agent's context — memory blocks, skills, prompts — lives as git-tracked files. So you can diff, review, branch, and roll back an agent's "mind" the same way you review code, instead of trusting an opaque vector store.
- Self-rewriting: agents edit their own system prompt and skills, and run periodic "dreaming" (sleep-time compute) to consolidate. Behavior is meant to drift toward your workflow over weeks rather than stay frozen at install time.
- One identity, many surfaces: the same agent is reachable from the CLI, a desktop app, the browser, and Slack/Telegram/Discord. Long-running, scheduled, or multi-agent work continues off your terminal instead of dying when you close the window.
Great Fit / Look Elsewhere
Great fit if you want a long-lived assistant that accumulates project knowledge over time, or you're exploring stateful and continual-learning agents — it comes from the team behind MemGPT. Look elsewhere if you mainly want a fast, predictable, stateless coding agent for one-off edits: persistence, dreaming, and self-rewriting add moving parts and unpredictability, and the harness is younger and more experimental than mainstream coding CLIs.
Where It Fits
Against stateless tools like Claude Code, Codex CLI, or Gemini CLI, which optimize for a clean blank-slate loop, Letta trades determinism for continuity. Reach for it when "the agent remembers" matters more than "the agent behaves identically on every run."
