Claude Subconscious
github.com/letta-ai/claude-subconscious- Category
- AI Agents
- Rank
- No. 430Tools index
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Coding
- Builder
- @letta-ai
- GitHub
- 2.9k stars
- Latest release
- v2.1.1
- Date
About
A background AI agent that gives Claude Code persistent memory by watching coding sessions, reading codebases, and whispering contextual guidance across projects. Built as a demo for the Letta Code SDK, it creates a subconscious layer that learns and remembers between Claude Code sessions.
What it does
Claude Code hooks feed completed session activity into a separate Letta agent. That agent can inspect project files, update shared memory, and return concise notes before prompts or tool use. Conversation mappings preserve continuity, while optional modes control whether Claude receives messages, full memory changes, or nothing.
Why it's ranked here
This is a compelling prototype because it joins session transcripts, code inspection, persistent memory, and timely context injection into one coherent loop. Read-only tools provide a sensible default, and self-hosted Letta is supported. The decisive caveat is explicit: its authors call it a demo and advise against production use.
What's good
The integration stays out of project instructions and injects context through hook output instead. Transcript delivery runs asynchronously after responses, so the memory agent should not block normal work. Shared memory can span projects, while separate agent identifiers allow isolation. Model discovery and fallback logic also reduce setup failures across Letta servers.
Tradeoffs
Every useful cycle depends on a Letta server, an API key, and a suitable model. Session transcripts and code context reach that agent, so deployment and tool permissions deserve scrutiny. Shared cross-project memory can mix unrelated context unless users assign separate agents. Full tool mode permits editing and shell access. The authors explicitly reject production use.
How to use it well
Use it as an experiment for developers who repeatedly return to related Claude Code projects and want preferences, architecture decisions, pending work, and recurring problems carried forward. Start with read-only tools and message-only injection, then isolate sensitive or unrelated projects with separate agents. It does not replace a production coding agent or production-grade memory service.
Technical notes+
package.json defines an ESM TypeScript package requiring Node.js 18 or newer, with @letta-ai/letta-code-sdk and tsx at runtime and Vitest for tests. scripts/session_start.ts creates or reuses a Letta conversation and stores session mappings. scripts/send_worker_sdk.ts resumes that conversation in a detached worker with configurable tool restrictions. scripts/sync_letta_memory.ts and scripts/pretool_sync.ts fetch assistant messages and memory changes for hook injection. scripts/transcript_utils.ts parses JSONL transcripts, truncates selected content, and formats XML. scripts/agent_config.ts resolves or imports an agent, validates identifiers, discovers models, and applies fallbacks. scripts/letta_api_url.ts normalizes hosted or self-hosted API endpoints.
Observed
- License
- MIT
- Primary language
- TypeScript
- Runtime
- Node.js 18 or newer, using ECMAScript modules
- Install surface
- Claude Code plugin marketplace commands or source installation with npm
- Interface
- Claude Code plugin implemented through session, prompt, tool-use, and response hooks
- Service support
- Hosted Letta API by default, with a configurable self-hosted Letta base URL
- Testing
- Vitest is configured, with colocated TypeScript tests for agent configuration
Read from README.md, package.json, scripts/agent_config.ts, scripts/pretool_sync.ts, scripts/letta_api_url.ts, scripts/session_start.ts, scripts/send_worker_sdk.ts, scripts/transcript_utils.ts, scripts/agent_config.test.ts, scripts/sync_letta_memory.ts, scripts/conversation_utils.ts.
What it can do
Monitor coding sessions in background
Active Claude Code session data → Session activity logs and metadata
Build persistent memory across sessions
Coding patterns, preferences, and session history → Accumulated knowledge base and user profile
Read and analyze codebases
Source code files and project structure → Codebase understanding and context mapping
Provide contextual coding guidance
Current coding context and historical memory → Relevant suggestions and recommendations
Remember project details between sessions
Project-specific information and code changes → Persistent project memory and context
Deliver whispered contextual hints
Real-time coding activity and stored context → Subtle guidance messages and suggestions
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.