AuthorsZhe Ren, Yibo Yang, Yimeng Chen et al.
TL;DR
GateMem uses long-form multi-principal episodes with hidden checkpoints to jointly score utility, access control, and active forgetting, revealing no method reaches high Memory Governance Score across all three.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
GateMem highlights that no evaluated method simultaneously achieves strong utility, robust access control, and reliable active forgetting across four domains and 2,218 checkpoints.
In hospital, office, campus, and household assistants, this means shared-memory agents can disclose protected information or recover deleted details, undermining institutional deployment and user trust.
HOW IT WORKS
GateMem centers on Episodes and Memory State, Checkpoints and Governance Categories, and the multiplicative Memory Governance Score (MGS) to capture governance behavior. GateMem also relies on LLM-as-a-judge evaluation and detailed attack-type breakdowns for access control and forgetting.
Conceptually, GateMem treats shared memory like a governed database with audit logs, not a simple chat history, where each query is checked against roles and deletion rules like access-controlled tables.
This design lets GateMem expose failures that a plain long context window cannot, such as post-deletion recovery under confirm yes no attacks or subtle cross-patient access violations.
DIAGRAM
This diagram shows how GateMem feeds an episode prefix and requester query into a shared-memory agent and then judges the response against hidden governance annotations.
DIAGRAM
This diagram illustrates how GateMem constructs domain scenarios, builds episodes, inserts checkpoints, and computes governance metrics.
PROCESS
Scenario Specification
GateMem defines Scenario Specifications Se with domain De, principals Pe, relationships Re, and initial access rules Ge to encode roles and scoped permissions.
Episode Construction
GateMem builds Episodes and Memory State Ee as turn sequences τt where facts, permissions, and deletion requests evolve and update M(e)_t via Ingest.
Checkpoint Construction
GateMem inserts Checkpoints and Governance Categories H with query_type qn, expected_action a⋆n, judge_spec Jn, and leak_targets Λn at selected turn boundaries.
Judge-based Evaluation
GateMem runs baseline agents on checkpoints and uses LLM-as-a-judge with structured specs to score Utility U, Access-Control Violation A, Forgetting Failure F, and Memory Governance Score.
KEY CONTRIBUTIONS
Formulating shared-memory governance
GateMem formulates multi-principal shared-memory evaluation as a governance problem where Utility, Access Control, and Active Forgetting are measured jointly via the Memory Governance Score.
Benchmark construction and protocol
GateMem introduces a benchmark with 91 episodes and 2,218 checkpoints using Episodes and Memory State, Checkpoints and Governance Categories, and LLM-as-a-judge with leak-target audits.
Baseline study of current agents
GateMem provides a baseline study across LONG-CONTEXT, RAG variants, A-MEM, MEM0, and REMEM, showing no system achieves high MGS while controlling both A and F.
RESULTS
Memory Governance Score
80.1%
+35.4 over RAG-NAIVE on Medical with GPT-5.4
Effective Utility U
91.4%
on Medical LONG-CONTEXT with GPT-5.4
Access-Control Violation A
10.4%
for LONG-CONTEXT vs 25.0% for RAG-NAIVE on Medical GPT-5.4
Active Forgetting Failure F
2.3%
for LONG-CONTEXT vs 7.9% for RAG-NAIVE on Medical GPT-5.4
GateMem evaluates shared-memory governance on Medical, Office, Education, and Household episodes with 2,218 checkpoints. The Medical GPT-5.4 results show LONG-CONTEXT reaches 80.1% MGS, while RAG-NAIVE only achieves 44.7%, quantifying the trade-off between utility and governance-aware memory handling.
BENCHMARK
GateMem evaluates shared-memory governance on Medical, Office, Education, and Household episodes with 2,218 checkpoints. The Medical GPT-5.4 results show LONG-CONTEXT reaches 80.1% MGS, while RAG-NAIVE only achieves 44.7%, quantifying the trade-off between utility and governance-aware memory handling.
BENCHMARK
Memory Governance Score (MGS) on GateMem Medical domain for GPT-5.4 backbones.
KEY INSIGHT
GateMem shows LONG-CONTEXT on Gemini-2.5-Flash-Lite reaches 93.9% utility on Education but still suffers 64.4% active-forgetting failures.
This is surprising because high utility usually suggests strong memory handling, yet GateMem reveals that excellent recall can coexist with frequent recovery of deleted information.
WHY IT MATTERS
GateMem gives researchers a concrete way to measure how shared assistants balance usefulness, access boundaries, and deletion compliance in multi-principal environments.
With GateMem, builders can design and tune memory systems specifically for institutional deployments, rather than relying on raw context length or recall-only benchmarks.
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…