<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> This project provides a persistent memory layer for the closed-source Claude Code agent, enabling it to retain context across sessions. By integrating with the Letta agent system, it allows Claude Code to accumulate patterns, learn user preferences, and receive asynchronous guidance, significantly enhancing its utility for developers and researchers working on long-term projects.

How It Works

The system operates by having a Letta agent observe Claude Code conversations. After each Claude Code response, the transcript is sent asynchronously to the Letta agent for processing. The agent then updates its memory blocks and prepares guidance, which is injected into the CLAUDE.md file before Claude Code's next prompt. This creates a feedback loop where the agent learns from interactions and provides context-aware suggestions, enhancing the main Claude Code agent's capabilities without direct modification of the closed-source model.

Quick Start & Requirements

Installation can be done via the plugin marketplace (/plugin marketplace add letta-ai/claude-subconscious) or by cloning the repository, running npm install, and enabling the plugin locally or globally. A mandatory LETTA_API_KEY obtained from app.letta.com is required. Optional environment variables include LETTA_AGENT_ID, LETTA_BASE_URL (for self-hosted Letta), and LETTA_MODEL for overriding the default LLM. Users on Linux systems with tmpfs may need to configure TMPDIR to avoid installation issues.

Highlighted Details

• Memory Architecture: Employs 8 distinct memory blocks for storing core directives, guidance, user preferences, project context, session patterns, pending items, self-improvement guidelines, and tool usage.
• Asynchronous Operations: Utilizes a fire-and-forget pattern with a detached background worker for sending transcripts to the Letta agent, preventing hook timeouts.
• Contextual Guidance: Provides guidance via <letta_message> blocks within CLAUDE.md, offering concise, observational, and non-intrusive suggestions.
• State Management: Manages durable state (conversations, session data) in .letta/claude/ and temporary log files in /tmp/letta-claude-sync/.
• Memory Diffing: Displays changes to memory blocks as diffs to Claude, optimizing token usage.
• Two-Way Communication: Allows Claude Code to directly address the Letta agent, fostering interactive dialogue.

Maintenance & Community

The provided README does not detail specific contributors, sponsorships, or community channels like Discord or Slack.

Licensing & Compatibility

The project is released under the MIT license, which is highly permissive and generally compatible with commercial use and integration into closed-source projects.

Limitations & Caveats

This project is explicitly labeled as experimental. It relies heavily on the closed-source nature of Claude Code, meaning its functionality is tied to that external system. The effectiveness of the subconscious agent improves over several sessions as it gathers more data, requiring an initial period of observation before providing significant value. Setup involves API key management and environment variable configuration, which may present a moderate barrier to entry.