RLM-Forge

What RLM-Forge does

The README states that RLM-Forge is not a new model architecture. Hermes Agent acts as the inner language model runtime, Ouroboros handles recursion, scheduling, state changes, termination and trace replay, and TraceGuard checks that the parent synthesis only claims facts backed by accepted child evidence handles. The design is inspired by the Recursive Language Models paper, arXiv 2512.24601, and the repository links its own paper on the approach.

The README reports a 24-cell live matrix across Hermes GLM, Claude Code Opus and Codex GPT-5.5 in which every cell passed the TraceGuard contract, and deterministic benchmarks on guarded memory that reject answer contamination. On a long-context truncation fixture it reports a tie with a single-call Hermes baseline. The project supplies local ouroboros and ooo wrappers, so uv run ouroboros rlm installs TraceGuard right after parent synthesis. Hermes MEMORY.md is used as a behavioral prior, never as factual evidence.

Key features

• Hermes Agent as the inner runtime for bounded JSON sub-calls
• Ouroboros outer scaffold for recursion, state and trace replay
• TraceGuard acceptance gate that rejects unsupported parent claims
• Memory priors from Hermes MEMORY.md used only for schema stability
• Benchmarks and experiment artifacts committed in the repository

When to use it

• Studying evidence-gated recursive execution on top of Hermes Agent
• Reproducing the repository's benchmarks on guarded agent memory
• Prototyping a recursive long-context workflow with auditable child evidence

Who it is for: Researchers and agent builders interested in recursive language model execution with verifiable evidence.

How it fits with Hermes Agent

Built around Hermes Agent as the inner runtime, and uses Hermes MEMORY.md as an operational prior.

Note: On its long-context test fixture the README reports a tie with a single-call Hermes baseline, so it presents the project as a runtime path rather than a quality gain.

FAQ