TL;DR for operators
An AI assistant can remember an accurate customer preference and still use it incorrectly. The relevant storage decision is whether that memory will improve appropriate future interactions without contaminating unrelated requests or crossing contextual boundaries.
ConMem evaluates candidate memories through three simulated future interactions, comparing responses with and without memory access. On its 1,079-case MemAlign benchmark, it achieves 75.16% admission accuracy and 65.31% Macro-F1, outperforming four evaluated baselines on both metrics. Yet it rejects only 39.33% of memories labeled Do-Not-Store. MemoryBank achieves higher balanced accuracy and substantially better rejection recall.
For teams operating persistent AI assistants, the contribution is a testable admission mechanism, not a demonstrated memory-safety solution. Separate checks for useful personalization, inappropriate retention, and contextual misuse should inform deployment decisions. The reported implementation averages 14.30 model-stage calls per candidate, making screening cost another material constraint.
An accurate memory can still be applied incorrectly
Suppose a customer tells a financial-services assistant that they prefer concise investment summaries. Retaining this preference could improve future reports. But the same information could produce inappropriate behavior if the assistant assumes the preference belongs to another customer, applies it outside its intended context, or treats a temporary instruction as permanent.
A memory’s factual accuracy does not settle whether it belongs in a persistent profile. Its value depends partly on the requests and circumstances in which the assistant will subsequently use it.
This creates a difficult admission problem. The system must decide whether to store information before knowing the future conversations in which it might appear.
In ConMem: Counterfactual Multi-Scenario Probing for Resolving LLM Memory Misalignment, Shuning Zhang and colleagues propose evaluating candidate memories through their simulated downstream behavioral consequences.1
The central finding deserves a qualified interpretation. Their method improves aggregate admission classification, but its ability to prevent inappropriate retention remains limited. On the primary benchmark, 182 of 300 memories labeled Do-Not-Store are nevertheless retained.
For an operator concerned about inappropriate personalization, that error rate may matter more than overall accuracy.
Test the future behavior before admitting the memory
ConMem changes the question from whether a candidate memory looks relevant or important to how its availability changes an assistant’s responses.
For every candidate memory, the framework generates three plausible future interactions:
- Relevant interaction: A situation in which the memory could provide useful personalization.
- Irrelevant interaction: A situation in which the memory should not influence the response.
- Risk-sensitive interaction: A situation that tests possible misattribution, outdated information, or inappropriate contextual carry-over.
Consider the preference for concise investment summaries. A subsequent request for an investment report might legitimately benefit from it. An unrelated technical question should remain unaffected. A request involving another customer’s preferences should not inherit the original customer’s profile.
ConMem generates a response to each simulated query without the candidate memory and another with the memory available. The model configuration is held constant, while the historical conversation is withheld from both response-generation conditions.
The historical context remains available to the evaluation components, which need it to determine whether personalization is justified.
Two distinct evaluators then examine the paired responses.
A blind response-quality judge assesses which response better serves the future request. It sees the historical context and query, but not the candidate memory or which response received it.
A separate memory-attribution auditor asks whether the candidate memory was substantively used, whether it materially changed the answer, and whether that change represents useful personalization, unnecessary personalization, neutral behavior, or risk.
This separation matters because a response can appear better without becoming better because of the memory. Likewise, two fluent responses may differ in whether they improperly carry user-specific information across contextual boundaries.
Each interaction contributes positive, negative, or neutral evidence. ConMem averages these signed contributions into a Counterfactual Memory Treatment Effect:
Here, $M$ is the proposed memory and $c_i$ represents the assessed contribution of one simulated interaction. Positive scores favor storage, negative scores favor rejection, and approximately zero scores trigger an additional evidence-based decision.
The mechanism therefore evaluates both the potential benefit of remembering something and the behavioral consequences of remembering it when it should not apply.
These are controlled comparisons within simulated interactions, not measured causal effects on actual users.
Higher admission accuracy does not mean stronger rejection
The authors construct MemAlign from previously user-annotated memory cases, retaining the original storage labels and feedback. They generate accompanying conversational histories and validate them through structural checks, semantic auditing, automated refinement, and researcher review.
After excluding two insufficiently grounded cases, the benchmark contains 1,079 candidate memories: 779 labeled Store and 300 labeled Do-Not-Store.
The primary benchmark comparison reports:
| Method | Accuracy | Macro-F1 | Balanced accuracy | Do-Not-Store recall |
|---|---|---|---|---|
| ConMem | 75.16% | 65.31% | 64.15% | 39.33% |
| A-MAC | 69.33% | 49.91% | 51.91% | 12.71% |
| MemoryBank | 67.19% | 64.30% | 67.95% | 69.67% |
Source: Section 5.3.1, Table 1. Values represent classification performance against annotated storage decisions.
ConMem improves accuracy by 5.83 percentage points and Macro-F1 by 15.40 points relative to A-MAC. Its rejection recall is also substantially higher than A-MAC’s.
But the comparison with MemoryBank changes the operational interpretation. MemoryBank rejects a considerably larger proportion of inappropriate memories and achieves higher balanced accuracy, despite lower overall accuracy and slightly lower Macro-F1.
The benchmark’s class distribution helps explain this result. Store cases account for 72.2% of observations. A system that performs well on retention decisions can achieve strong overall accuracy even when it frequently admits memories that should be rejected.
ConMem’s Store recall is 88.96%, compared with only 39.33% Do-Not-Store recall.
For a customer-service operator, the preferred metric therefore depends on the relative consequences of rejecting a useful memory and retaining an inappropriate one. The paper measures agreement with storage preferences, not the monetary, privacy, or customer-experience cost of either error.
Attribution auditing contributes more than response preference alone
The paper’s component ablations help explain which parts of the mechanism contribute to its benchmark performance.
Removing the blind quality judge reduces Macro-F1 from 65.31% to 59.87%. Removing the memory-attribution auditor produces a larger decline, to 52.41%.
These are ablation results on the same benchmark, rather than independent demonstrations of real-world safety. Nevertheless, they indicate that response preference alone provides insufficient information for this admission policy.
The attribution auditor contributes information about whether the proposed memory actually influenced the response and whether that influence was appropriate.
The treatment of inconclusive evidence also matters. The primary evaluation produces 199 approximately zero-score cases, representing 18.44% of candidates. Replacing adaptive arbitration with a fixed Store decision lowers Macro-F1 to 56.83%; always rejecting these cases produces 60.38%.
For agent-quality testing, these findings suggest keeping response-quality assessment, memory-use attribution, and inconclusive-case handling as distinct functions rather than collapsing them into a single preference judgment.
Contextual boundary violations remain the weakest cases
ConMem’s overall rejection recall conceals substantial differences among error categories.
The paper’s category-level error analysis reports rejection recall of just 12.90% for context_misapplication, 16.67% for incomplete_memory_use, and 21.21% for granularity_mismatch.
By comparison, rejection recall reaches 57.41% for false_memory_use and 58.82% for stale_memory_use.
These results suggest that the framework is more effective at identifying certain false or outdated memory uses than subtle contextual scope violations.
The distinction matters in enterprise personalization. An assistant might correctly recall a preference but apply it to the wrong person, transaction, or task. Such failures are not necessarily detectable through factual-memory checks alone.
Further evaluations provide useful boundaries rather than eliminating these concerns. On a balanced 400-case HaluMem subset, ConMem achieves 66.00% accuracy while rejecting only 38.50% of interfering memories.
Backbone comparisons report higher Macro-F1 with GPT-5.6-sol and DeepSeek-V4-Flash than with the primary Qwen3.8-Max configuration. GPT-5.6-sol also improves Do-Not-Store recall to 55.33%, although substantial inappropriate retention remains.
Across 27 predefined scoring configurations, the primary backbone’s accuracy ranges from 75.16% to 78.13%. Changes in risk penalties shift the balance between retaining appropriate memories and rejecting inappropriate ones.
These are sensitivity and robustness evaluations, not evidence that the framework generalizes reliably to all languages, tasks, or production environments.
Memory admission becomes a measurable operating cost
ConMem’s additional screening requires multiple model operations for each memory candidate: generating and validating scenarios, producing paired responses, evaluating quality, auditing attribution, and sometimes resolving inconclusive scores.
The primary experiment reports 15,430 model-stage calls across 1,079 candidates, averaging 14.30 calls per candidate.
Mean summed successful API latency is 347.64 seconds per candidate. This figure is not end-to-end wall-clock time: the experiment permits concurrent processing, and the reported statistic excludes failed attempts and retry-backoff delays. Token-level costs are not supplied.
For operators considering the approach, three measurements would be particularly relevant: the rate of inappropriate memories admitted, the rate of useful memories rejected, and the inference expense of screening each candidate.
A possible business implementation would apply deeper prospective testing to memories with material cross-session consequences, while evaluating simpler admission policies for less consequential information. That is a Cognaptus deployment inference, not a strategy tested by the authors.
The remaining validation gap is substantial. MemAlign uses synthetic histories reconstructed from annotated cases, and future interactions are model-generated. Both quality and attribution depend on LLM evaluators. The original storage labels reflect user annotations rather than observed downstream outcomes. The source also contains internal numerical inconsistencies, although its main benchmark table provides the comparative figures reported here.
Accordingly, the results establish a promising computational evaluation method, not reliable prevention of privacy violations, misattribution, or contextual misuse in deployed agents.
The storage decision should anticipate inappropriate use
ConMem provides a concrete way to examine an otherwise neglected part of persistent-agent design: what storing a memory might cause the assistant to do later.
Its most transferable contribution is the admission procedure itself. A candidate should demonstrate benefit when relevant, limited influence when irrelevant, and appropriate behavior when contextual boundaries change.
Yet the measured rejection failures show why improved aggregate accuracy cannot be treated as sufficient memory governance.
For teams building persistent assistants, counterfactual probing offers a stronger basis for evaluating storage decisions than relevance scoring alone. Its practical value will depend on whether future implementations can identify difficult boundary violations more reliably and deliver that screening at an acceptable operating cost.
The central requirement remains precise: a memory should be retained only when its anticipated benefits justify the risk of its subsequent use.
Cognaptus: Automate the Present, Incubate the Future.
-
Shuning Zhang and Tenghao Ji and Jingruo Chen and Eryue Xu and Bohao Chu and Xin Yi (2026). ConMem: Counterfactual Multi-Scenario Probing for Resolving LLM Memory Misalignment. arXiv:2610.33432. https://arxiv.org/abs/2610.33432 ↩︎