Revocable Learned State Enhancements via Process Sidecars Introduced by John Sweeney

On June 29, 2026, John Sweeney presented a new paper titled Revocable Learned State via Process Sidecars, which explores advancements in machine learning memory adaptation. The work introduces a novel approach to revoking learned states in language models, aiming to improve safety and reliability in AI outputs.

Understanding Process Sidecars in Machine Learning

The concept of process sidecars involves a two-coefficient edit family defined as $\theta(\lambda,\gamma)=\theta_{\mathrm{AMS}}-\lambda\Delta_{\mathrm{M}}-\gamma\hat{R}_{\mathrm{S}\leftarrow\mathrm{M}}$. This methodology addresses the challenges of memory direction transportation in AI systems. By utilizing a centered secant through the realized future AdamW safety-training process, this approach enhances the precision of memory edits.

The implementation of process sidecars utilizes $\varepsilon=1$ at the natural memory-edit scale, allowing for the reuse of $\theta_{\mathrm{AMS}}$ as the positive endpoint. This innovative method computes an additional safety trace at $\theta_{\mathrm{A}}-\Delta_{\mathrm{M}}$, demonstrating significant improvements in AI safety.

Key Findings from the Research

Sweeney's research yields two primary findings. First, the exact sidecar, utilizing the true transported direction $R_{\mathrm{S}\leftarrow\mathrm{M}}$ at $(\lambda,\gamma)=(1,1)$, successfully recovers the counterfactual safety-only oracle $\theta_{\mathrm{AS}}$ up to second order. This proof treats AdamW as an augmented-state map over parameters, first moments, and second moments.

Second, the necessity of process information is highlighted; any future safety training that alters memory direction results in first-order counterfactual errors. The process-sidecar edit is shown to be second-order accurate, making it a crucial enhancement over traditional methods.

Improving AI Safety Through Enhanced Memory Management

Across three models tested, the validation-selected 2D edit demonstrated improved held-out refusal closure compared to naive task arithmetic. In all trials, the process-JVP subfamily, particularly the diagonal slice of the cached 2D grid, showed superior performance.

This research opens new avenues for enhancing the safety and reliability of AI systems by refining how learned states are managed and revoked, ultimately leading to more robust language models.