Jacobian-Causal Memory Control for Persistent Memory in Language Model Systems
Summary
- A new preprint introduces Jacobian-causal memory control (JCMC) for managing persistent memory in language model systems.
- JCMC represents memory interventions via predicted influence signatures, selects interventions against targets and side-effect budgets, and checks effects before retaining or rolling back.
- The paper provides formal proofs including a quadratic local-approximation bound, selection regret bounds under signature error, and non-additive interaction bounds.
- The work includes no empirical benchmark measurements, instead defining a paired counterfactual estimator and evaluation protocol comparing JCMC against baseline retrieval methods.
Why it matters
This research provides a formal theoretical framework for controlling persistent memory interventions in language models, helping to avoid unintended side effects during inference.