lambda-symbolics/autolith

What it solves

Autolith is a terminal-based AI agent that operates within a live Common Lisp image. It solves the problem of rigid agent architectures by allowing the agent to observe, introspect, and modify its own code and state in real-time, enabling it to adapt to user preferences and specific task requirements without losing stability.

How it works

Autolith integrates a Common Lisp REPL as its primary prompt input, allowing users to mix prose prompts and exact Lisp forms. It utilizes a live image that can be snapshotted and rolled back to recovery images to prevent permanent failure during self-modification. For handling large inputs, it uses "recursive inference" (RLM), which treats large files as environments to slice, search, and sub-infer over them using isolated Lisp REPLs rather than stuffing everything into a single prompt context window.

Who it’s for

It is designed for developers and power users who are comfortable with Lisp and prefer a terminal-centric workflow, seeking an agent that can be dynamically modified and self-evolving.

Highlights

• Lisp-first control: The prompt window is a full REPL where users can execute Lisp forms to compute prompts or call tools directly.
• Self-modification: The agent can perform "live brain surgery" on itself with built-in tools for inspection and modification, supported by a system of checkpoints and recovery images.
• Recursive Inference (RLM): Handles massive inputs by breaking them down into isolated frames of inference with shared budgets and validated data returns.
• State Recovery: Uses append-only files and a conversation-scoped vault to ensure steering messages and follow-ups survive crashes.
• Cross-platform: Supports Linux, MacOS, and various BSDs.