Letta gives Claude Code a literal subconscious

Claude Subconscious is a new open-source plugin from Letta that basically gives Claude Code a split brain.

Instead of wiping your context clean every time you restart a terminal session, the plugin runs a persistent secondary Letta agent right underneath your main interface. While you work, this background agent quietly watches the session, reads your repo, and builds long-term memory over time.

What it actually does

This background helper is far more than a passive scratchpad for notes.

While scanning your session transcripts, it actively navigates your files using built-in Read, Grep, and Glob tools. It digs through the codebase on its own to build an actual understanding of how your project is structured.

All of this knowledge is organized into 8 persistent memory blocks. These blocks track everything that usually gets lost between sessions, including your personal coding preferences, project architecture, pending TODOs, and recurring design patterns.

Even better, the memory is globally shared across all your projects. A single subconscious agent manages separate conversation threads for each repository, meaning lessons and habits learned in one codebase naturally carry over to the next.

How the split brain works

Speed is critical when you are in the flow, so the plugin is designed to be completely non-blocking.

Transcript processing runs asynchronously via the Letta Code SDK using background hooks, so Claude Code never stutters or hangs while you type. When the background agent has relevant context ready, it silently whispers the guidance directly into Claude Code's context via stdout before each new prompt, without writing junk files to your disk.

Setup is automatic on first launch, and it supports multiple operational modes. In addition to Anthropic, it can route background processing through models from OpenAI, Google, and other major providers.

Why it matters

CLI coding agents are great at execution, but their total lack of long-term memory makes multi-day projects frustrating. Letta's approach is clever because it decouples memory from execution: the main agent focuses purely on writing code, while the background agent handles repo awareness and context retrieval without getting in the way.

If you use Claude Code every day, this is an easy recommendation to try right now. Just keep in mind that running a secondary agent with continuous repo searches and background calls will add some extra token consumption to your API bill.