"GraphMemShield: Auditing Cross-Session Leakage in Graph Memories," by Phat T. Tran-Truong, Dr. Xuan-Bach Le and Son Ha, has been accepted at CIKM 2026 (ACM International Conference on Information and Knowledge Management), Full Research Track, CORE Rank A, with an acceptance rate of 597/2216 (27%). The conference takes place in Rome, Italy, in November 2026. Knowledge-graph memories are becoming a practical substrate for AI assistants: a system extracts entities, relations, timestamps and provenance from one session, then retrieves k-hop neighbourhoods to ground an answer in another. That creates a privacy boundary unlike either ordinary model memorisation or document-RAG leakage — the sensitive object here is a live, multi-user, provenance-tagged graph. GraphMemShield packages a set of scoped retrieval-risk invariants that bound cross-session exposure, alongside a separate local-DP statement for a fixed-universe randomized-response mechanism. It implements attacks for cross-session edge exposure, adaptive probing, session linkage, temporal inference and response-content leakage. On a 2,304-edge enterprise benchmark, unguarded retrieval leaks 11.8±0.6 edges and 7.9±0.7 sensitive terms. Bounded sharing makes the trade-off explicit: raising the per-pair budget from 0 to 10 lifts downstream graph-QA accuracy from 0.00 to 0.71, whereas strict isolation yields 0.00. The authors also show that the guarantee depends on correct sensitivity provenance, and propose three mitigation recipes that restore most of the bound even when half the labels are corrupted.