Whose Memory Is It? Scope-Aware Commit Rules for Long-Term LLM Memory

Reasoning systems are trained to imagine alternatives, they branch over candidate solutions, write plans, revise beliefs, and reflect on failed attempts. Persistent memory creates a new failure mode. Once these useful thoughts outlive the episode that produced them, a hypothesis or counterfactual can return as a fact. The first surprise is that the language model already carries the distinction that memory needs. Across two Qwen2.5 scales, controlled interventions show that hidden representations can route the same proposition to the appropriate discourse owner; the distinction is lost when a memory interface stores content but discards its linguistic scope. We introduce $\text{Cask}$ (Causally Anchored Scoping Keys), which turns that internal distinction into a typed actual/non-actual memory edge. The second surprise is that directly storing the localized coordinates is unstable: equally valid bases rotate those coordinates and silently change the external key. $\text{Cask}$ instead stores the relations among diagnostic responses, yielding a compact interface that is invariant to this arbitrary choice. We prove invariance, completeness, and a margin-based stability condition. On source-isolated long-conversation conflicts, $\text{Cask}$ is the strongest evaluated attachment method at both model scales, reaching 83.33\% on the 3B model, and it reduces cross-space contamination while preserving answer quality. The result connects two rapidly advancing directions: richer reasoning produces more provisional worlds, and durable memory needs a principled commit operation that keeps those worlds apart.