AuthorsJong Wook Kim, Byoungjae Min, Kennedy Edemacu et al.
TL;DR
DP-MemView privately selects public memory views with per-attribute DP ledgers, keeping transcript AUC near 0.49 while preserving personalization quality around 0.88.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Long-term memory enables persistent personalization, but repeated memory-conditioned responses can cumulatively reveal protected attributes even when never stated explicitly.
In a meal-planning assistant, ordinary queries yield blood-sugar-conscious advice that lets a semi-trusted application infer type-2 diabetes from the observed transcript.
HOW IT WORKS
DP-MemView centers on Slot-Level View Selection, Attribute Accounting, and the DP-MemView Interface Contract to privatize memory-conditioned views.
Think of DP-MemView as a privacy-aware memory controller: raw slots stay like protected disk blocks, while public views act like safe RAM hints for the LLM.
This design lets DP-MemView bound full adaptive transcripts with pure Ba-DP, something a plain context window or raw memory retrieval cannot guarantee.
DIAGRAM
This diagram shows how DP-MemView processes each turn of an adaptive application, from query to differentially private transcript entry.
DIAGRAM
This diagram shows how DP-MemView is evaluated on PairedMem and public-corpus tracks, including baseline interfaces and safeguard ablations.
PROCESS
Slot-Level View Selection
DP-MemView uses Slot(qt, ˜τ<t) to choose active slots Gt, then applies the read policy Rt,g to map queries onto memory positions.
Attribute Accounting
DP-MemView computes Γ⋆(Rt,g) and declares Γt,g, ensuring every intersecting protected group Ia is charged in the per-attribute ledgers La.
Exponential Mechanism View Sampling
Using the internal scorer ut,g and sensitivity bound ∆ut,g, DP-MemView samples a view vt,g from Vg with the exponential mechanism and local εt,g.
Interface Contract and Transcript Release
Under the DP-MemView Interface Contract, the response LLM KLLM receives only qt, vt, and τ<t, producing yt that composes into pure Ba-DP for ˜τT.
KEY CONTRIBUTIONS
Adaptive Transcript Privacy Formulation
DP-MemView formalizes long-term memory privacy as pure Ba-DP over full adaptive transcripts ˜τT, not single responses, with attribute-level adjacency M ∼a M′.
DP-MemView Interface Contract
DP-MemView defines Conditions C1–C6, including fixed vocabularies Vg, complete charging Γt,g ⊇ Γ⋆(Rt,g), and pathwise caps La ≤ Ba with generic fallbacks.
Empirical Evaluation With Safeguard Ablations
DP-MemView shows AUC near 0.49 and U around 0.88 across three LLMs, while ablations like CappedArgmax push AUC to 0.754 and break the Ba-DP bound.
RESULTS
AUC
0.490
-0.352 vs RawReadSet on PairedMem with Qwen2.5-7B
TPR@5
0.016
-0.531 vs RawReadSet on PairedMem with Qwen2.5-7B
tRec
0.506
-0.015 vs RawReadSet on PairedMem with Qwen2.5-7B
U
0.877
+0.102 over RawReadSet on PairedMem with Qwen2.5-7B
On the PairedMem synthetic benchmark at T = 16 with Qwen2.5-7B, which tests attribute distinguishability under adjacent memory stores, DP-MemView (on) sharply reduces AUC while keeping overall utility U comparable to GenericOnly. This MAIN_RESULT shows DP-MemView can maintain near-chance transcript privacy without sacrificing personalization quality.
BENCHMARK
On the PairedMem synthetic benchmark at T = 16 with Qwen2.5-7B, which tests attribute distinguishability under adjacent memory stores, DP-MemView (on) sharply reduces AUC while keeping overall utility U comparable to GenericOnly. This MAIN_RESULT shows DP-MemView can maintain near-chance transcript privacy without sacrificing personalization quality.
BENCHMARK
AUC for different memory interfaces on PairedMem synthetic, measuring transcript-level attribute distinguishability.
BENCHMARK
AUC for DP-MemView and baselines on public-corpus transfer, using corpus-derived memory text.
KEY INSIGHT
DP-MemView (pre) achieves tRec up to 0.571 on PairedMem with Gemma-2-9B while keeping AUC as low as 0.464.
This is surprising because stronger personalization usually increases leakage, yet DP-MemView’s interface contract lets higher tRec coexist with near-chance attribute distinguishability.
WHY IT MATTERS
DP-MemView unlocks attribute-level differential privacy for entire adaptive transcripts, not just single answers or masked spans.
Builders can now connect long-term memory agents to semi-trusted applications while bounding how much transcripts can change adversaries’ prior odds about protected attributes.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…