Zening Li, Hongchao Qin, Guoyuan Zeng, Rong-Hua Li · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840874
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Long-lived AI agents increasingly store private memory graphs of user goals, retrieved documents, tool calls, failures, and repairs. Mining these memories across users is crucial for discovering recurrent operational patterns, yet the problem remains unsolved: raw graphs cannot be centralized, and even aggregate motifs can reveal workflows supported by only one or two users. Existing local differential privacy (LDP) frequency-estimation and pattern-mining methods optimize top-K recovery, but they have no principled release-time defense against rare-workflow disclosure. We argue that the missing axis is not a stronger randomizer but a safer server-side release policy. We present PrivMem, which extracts retrieval-causal temporal motifs on device, perturbs them with per-coordinate randomized response, and releases them through a shrinkage ranker, a public-schema risk gate, and a Benjamini--Hochberg calibrated false-discovery filter. The analysis proves per-coordinate ε-LDP, derives concentration for the unbiased support estimator, establishes a support-gap condition for top-K recovery, and shows that the release stack preserves privacy by post-processing. On a 240-user planted benchmark, PrivMem reduces in-schema rare-workflow leakage by over 35× at matched utility; at ε=1,0=8, it keeps top-20 F1 at 0.86 while cutting leakage from 0.27 to 0.007. It also drives out-of-schema leakage to zero and improves released-subspace clustering by over 10× ARI at ε=2. On a 1, 757-user real-trace corpus across eight LLM-agent domains, PrivMem releases exactly zero rare motifs at every threshold, while LDP baselines leak 0.28-0.57.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.