LLMs living in a persistent Common Lisp REPL.
Rhema (Greek ῥῆμα — the spoken, performative word) is a system that embeds large language models inside a persistent Common Lisp REPL. The model doesn't generate Lisp on demand and discard it — it lives in the REPL. It defines its own tools, evolves them across sessions, and uses the REPL as external memory and computation.
- Homoiconicity — code is data. The gap between what a model thinks and what it writes is minimal.
- Uniform s-expression syntax — a trivial grammar means fewer generation errors and simpler parsing.
- Token efficiency — more semantic content per token than most languages.
- Frozen spec — Common Lisp was standardized in 1994 and hasn't changed. Every model's training data is consistent.
The conceptual framework comes from de la Torre (2025):
"From Tool Calling to Symbolic Thinking: LLMs in a Persistent Lisp Metaprogramming Loop" — arXiv 2506.10021v1
The paper proposes embedding LLMs in a persistent Lisp metaprogramming environment but provides no implementation. Rhema is that implementation.
Alonzo Church (a Presbyterian) gave us lambda calculus — the foundation of functional programming. Turing gave us mutable state — the foundation of imperative programming. One might argue that functional languages are, therefore, the Lord's language. We leave the theological implications as an exercise for the reader.
Early development. Repo bootstrap in progress.