AuthorsWenjie Wang, Wenhe Si, Xinyue Xu, Yue Xu
TL;DR
SP-Mem decouples sanitized memory from protected private values, enabling privacy-aware retrieval while keeping privacy-entity identification at 0.996 accuracy and unnecessary exposure as low as 0.33%.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
SP-Mem targets memory systems that optimize utility but neglect privacy, even though privacy-entity identification can reach 0.996 accuracy with careful design.
In dinner recommendation or form-filling tasks, long-term conversational agents may retrieve and expose home addresses or other PII, creating persistent privacy risk across sessions.
HOW IT WORKS
SP-Mem combines Privacy-Aware Memory Writing, Partitioned Storage with Privacy Mapping, and Privacy-Aware Query-Time Reasoning and Authorized Retrieval over complementary vector and graph memory.
You can think of SP-Mem like a library card catalog: sanitized entries stay searchable on open shelves, while exact private values are locked in a vault linked only by keys.
This design lets SP-Mem retrieve rich personalized context beyond a plain context window while restoring exact private values only when tasks require them and users grant consent.
DIAGRAM
This diagram shows how SP-Mem analyzes a query, decides if private values are needed, and conditionally hydrates them under user consent.
DIAGRAM
This diagram shows how SP-Mem's benchmark generates user profiles, subtasks, history dialogues, and test queries.
PROCESS
Privacy-Aware Memory Writing
SP-Mem uses Privacy-Aware Memory Writing to convert user utterances into facts and relation triplets while tagging private values with the privacy-entity taxonomy.
Partitioned Storage with Privacy Mapping
SP-Mem applies Partitioned Storage with Privacy Mapping to store sanitized facts in vector and graph stores and exact private values in a protected private store linked by mapping keys.
Privacy-Aware Query-Time Reasoning and Authorized Retrieval
SP-Mem runs Privacy-Aware Query-Time Reasoning and Authorized Retrieval to decide if private entities are required, request consent, and selectively hydrate exact values via the privacy mapping layer.
Response Generation
SP-Mem performs response generation using the current query, sanitized memories, and any authorized exact private values to complete tasks while respecting privacy constraints.
KEY CONTRIBUTIONS
Sanitized Privacy-Mapped Memory Architecture
SP-Mem introduces a Sanitized Privacy-Mapped Memory architecture combining Privacy-Aware Memory Writing, Partitioned Storage with Privacy Mapping, and Privacy-Aware Query-Time Reasoning and Authorized Retrieval over hybrid vector plus graph memory.
Privacy-Aware Memory Benchmark
SP-Mem comes with a privacy-aware memory benchmark of 1,000 user profiles, 21,000 history dialogues, and 54,000 user-query evaluation instances across finance, medical, education, and mental support domains.
Joint Evaluation of Utility and Privacy
SP-Mem provides an evaluation pipeline that measures response quality, personalization, privacy behavior, and token cost, including metrics like Unnecessary Privacy Usage dropping to 0.33% on preference-only tasks.
RESULTS
Privacy-entity identification accuracy
0.996
overall accuracy across domains for SP-Mem privacy-aware memory writing
Privacy-entity identification recall
0.992
recall across domains for SP-Mem privacy-aware memory writing
Privacy-entity identification precision
0.965
precision across domains for SP-Mem privacy-aware memory writing
Preference-only UPU rate
0.33%
exact private exposure on preference-only tasks for SP-Mem vs 16.00% for Full-context
The privacy-aware benchmark tests SP-Mem on 2,100 history dialogues and 5,400 queries, measuring privacy-entity identification and Unnecessary Privacy Usage. These results show that SP-Mem accurately detects private entities while sharply reducing unnecessary exposure compared to Full-context prompting.
BENCHMARK
The privacy-aware benchmark tests SP-Mem on 2,100 history dialogues and 5,400 queries, measuring privacy-entity identification and Unnecessary Privacy Usage. These results show that SP-Mem accurately detects private entities while sharply reducing unnecessary exposure compared to Full-context prompting.
BENCHMARK
Unnecessary Privacy Usage rate when tasks require only preferences and no private information.
KEY INSIGHT
SP-Mem keeps Unnecessary Privacy Usage on preference-only tasks at just 0.33%, while Full-context prompting leaks exact private values in 16.00% of responses.
This is surprising because many assume more context always helps personalization, but SP-Mem shows that trimming exact private values can improve safety without hurting task completion.
WHY IT MATTERS
SP-Mem unlocks long-term conversational agents that remember rich user preferences while tightly controlling when exact private values are stored, retrieved, and revealed.
Builders can now deploy memory-augmented assistants that support tasks like finance and medical help with explicit consent gates, rather than relying on brittle prompt-time sanitization.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…