Letta Code desktop tutorial: memory and agent setup

The speaker introduces the Letta Code desktop app, which he describes as available for macOS, Windows, and Linux. The speaker explains its agent-centered interface: each agent has memory that persists across conversations, with a main chat for ongoing work and separate threads for parallel tasks.

The memory walkthrough covers MemFS, which the speaker describes as Git-backed and versioned. Core memory stays in the system prompt; external memory loads when needed. A graph view displays links between memories, while history and diffs expose edits. The demonstration forks a conversation to organize several desktop folders in parallel. Default permissions pause file operations for approval, and the speaker switches to unrestricted mode for the demo.

Agents execute on the user's machine by default. The speaker also describes remote execution through the Letta Code NPM package. Local execution is distinct from model inference: the app connects to API keys, coding plans, or inference through a Letta account.

For AI agent setup, /init can draw on Claude Code or Codex conversation history. Dream agents review recent messages and update memory; the speaker recommends a trigger of roughly 20 steps or more rather than the two-step demo setting. The /doctor skill reorganizes bloated memory. The tutorial also covers deploying another existing agent as a subagent, importing skills, and configuring schedules and messaging channels.