Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM Agents

TL;DR - This paper formalizes execution-state unlearning for stateful LLM agents and proposes provenance-guided selective replay to remove revoked information from prompts, memory, plans, and KV caches. The method matches a full reset’s behavior while recomputing up to 9× fewer tokens.

• Exact unlearning requires recomputing at least (T-\tau+1) transitions after the target information’s injection step.
• Selective replay restores a pre-injection checkpoint, crops the KV cache, and regenerates the affected suffix from sanitized state.
• Deleting memory records alone did not reduce leakage, while instruction-based forgetting reached Leak@probes = 1.00 under elicitation.
• Source redaction still applied a revoked preference in 80% of episodes; selective replay was indistinguishable from a full reset in the reported audits.