Autolith: A self-modifiable general purpose Lisp AI agent

Autolith is a self-modifying, general-purpose AI agent implemented as a live Common Lisp image that you interact with through a terminal REPL. It can inspect and alter its own code and runtime state, create immutable generations and checkpoints, and recover by reloading a pristine recovery image when needed. The project emphasizes ergonomics and unobtrusive terminal visuals, follows XDG-like standards, and aims for a Lisp-machine feel rather than a toy. It supports Linux, macOS, multiple BSDs and Windows (from source), ships standalone binaries and a Nix runtime, and offers a simple installer and provider authentication for many model backends.

Functionally, Autolith exposes Lisp-first control: ordinary messages are just prompts and exact Lisp forms can be evaluated in the active image or in isolated scratch REPLs. There are dedicated self-inspection and modification tools, ephemeral worker REPLs, and a recursive inference system (rlm.*) that treats large inputs as environments, fans out bounded sub-inferences under shared budgets, and persists traces for auditability. Conversation state is stored append-only in a vault with commands to list, restore, or discard, and the agent supports updates and rollbacks. The code is ISC licensed and integrates with related Lisp projects and tooling.