AuthorsHanxiang Chao, Yihan Bai, Rui Sheng et al.
TL;DR
STALE uses a three-dimensional probing framework over 400 implicit-conflict scenarios to show CUPMEM’s write-side state adjudication reaches 68.0% overall accuracy where other systems stay below 55.2%.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
STALE shows that even the best evaluated model, Gemini-3.1-pro, reaches only 55.2% overall accuracy on implicit-conflict resolution.
In realistic long-term personalization, this means LLM agents keep stale memories, accept false premises, and produce unsafe or incoherent assistance.
HOW IT WORKS
STALE structures evaluation around State Resolution, Premise Resistance, and Implicit Policy Adaptation, while CUPMEM adds write-side belief updating and topology-triggered belief propagation.
You can think of STALE as a stress test for a user’s evolving “state RAM,” and CUPMEM as a card catalog that rewrites or retires cards whenever new evidence arrives.
This explicit adjudication lets CUPMEM retire outdated beliefs before retrieval, something a plain context window or naive RAG stack cannot enforce.
DIAGRAM
This diagram shows how STALE probes State Resolution, Premise Resistance, and Implicit Policy Adaptation over Type I and Type II conflicts.
DIAGRAM
This diagram shows how STALE constructs conflict scenarios from base states, adversarial conflicts, quality control, and haystack packaging.
PROCESS
Base State Formulation
STALE samples a user attribute from the topic ontology and uses Base State Formulation to generate an earlier observation mo anchoring a clear belief.
Adversarial Conflict Generation
Using Adversarial Conflict Generation, a Logic Attacker writes a later observation mn that creates Type I or Type II implicit conflict without explicit negation.
Quality Control
In Quality Control, an LLM judge filters candidate pairs, enforcing Belief Incompatibility and Non explicit Invalidation before STALE accepts the scenario.
Multi turn Dialogue Packaging
STALE wraps mo and mn into multi turn sessions, inserts them into a long context haystack with distractors, and attaches SR, PR, and IPA probes.
KEY CONTRIBUTIONS
Latent user state tracking
STALE formalizes conversational memory as latent user state tracking with implicit conflict defined by Belief Incompatibility and Non explicit Invalidation, across 400 scenarios and 1,200 queries.
Three dimensional probing framework
STALE introduces State Resolution, Premise Resistance, and Implicit Policy Adaptation probes, revealing that Gemini 3.1 pro reaches only 55.2% overall accuracy.
CUPMEM prototype for state adjudication
CUPMEM adds write side belief updating and propagation aware search, boosting GPT 4o mini from 8.7% to 68.0% overall accuracy on STALE.
RESULTS
Overall
68.0%
+12.8 over Gemini-3.1-pro
Type I PR
78.0%
vs GPT-4o-mini backbone at 0.0%
Type II PR
75.0%
premise resistance gain over Gemini-3.1-pro at 14.0%
Type I SR
91.0%
close to Gemini-3.1-pro at 92.0% with smaller backbone
STALE evaluates implicit conflict handling over 400 scenarios and 1,200 queries, measuring SR, PR, and IPA for Type I and Type II conflicts. CUPMEM’s 68.0% overall accuracy shows that explicit state adjudication can close much of the gap left by Gemini-3.1-pro’s 55.2% on the same benchmark.
BENCHMARK
STALE evaluates implicit conflict handling over 400 scenarios and 1,200 queries, measuring SR, PR, and IPA for Type I and Type II conflicts. CUPMEM’s 68.0% overall accuracy shows that explicit state adjudication can close much of the gap left by Gemini-3.1-pro’s 55.2% on the same benchmark.
BENCHMARK
Overall accuracy on STALE across closed source LLMs, open source LLMs, memory frameworks, and CUPMEM.
KEY INSIGHT
STALE shows that Gemini 3.1 pro hits 92.0% on Type I State Resolution but only 30.0% on Type I Premise Resistance.
This breaks the assumption that once a system can recognize stale information explicitly, it will automatically reject user queries that embed the same stale premise.
WHY IT MATTERS
STALE and CUPMEM make it possible to test and enforce when LLM agents should retire or revise long term personalized memories.
Builders can now design agents that resist stale premises, propagate updates across related attributes, and maintain coherent user state over 150K token histories.
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…