CrossMemBench

A Memory Benchmark for Cross-Domain Preference Transfer and Conflict Preservation

Benchmark Overview · Dataset · Tasks & Metrics · Citation

Benchmark Overview

What is cross-domain personalization? Users express preferences in domain-specific contexts, but the underlying preference may remain relevant to decisions in other areas of their lives, while preferences from different domains may also impose competing constraints on the same decision.

Key Points

• Two complementary evaluation settings: CMRT evaluates whether agents can transfer latent user preferences across domains, while DCA evaluates whether they can jointly preserve conflicting preferences expressed in different domains.
• Five functionally distinct life domains: education, travel, health, finance, and food/dining.
• 40 coherent user profiles and 560 benchmark items: including 200 CMRT and 360 DCA questions with controlled ground-truth memories.
• Evaluation under memory interference: systems are tested with 0, 100, 500, and 1,000 distractor memories to measure retrieval robustness as memory stores grow.
• Memory-grounded scoring: evaluation requires both a correct decision and a justification grounded in the relevant memories, while retrieval is tracked separately from downstream task accuracy.

Overview of CrossMemBench. Left: CMRT tests whether an agent can abstract a preference from one life domain and apply it in another. Right: DCA tests whether an agent can retrieve opposing cross-domain preferences and recommend a response that preserves both constraints.

Dataset

The dataset is organized as follows:

data/
├── profiles.json
├── noise.json
├── u001/
│   └── u001.json
├── u002/
│   └── u002.json
└── ...

• profiles.json contains the descriptions for all 40 user profiles.
• noise.json contains approximately 2,200 distractor memories used to evaluate retrieval robustness under increasing memory interference.
• uXXX/uXXX.json contains the benchmark data associated with each individual user profile.

We release the prompts used during dataset construction in the CrossMemBench GitHub repository. The dataset generation and human validation procedures, along with associated design choices, are described in detail in the paper.

The evaluation code is available in the CrossMemBench GitHub repository.

Tasks & Metrics

CrossMemBench evaluates whether memory-augmented LLM agents can transfer latent user preferences across domains and preserve conflicting cross-domain constraints in personalized decision-making.

• CMRT: Measures whether a memory expressed in one domain can be surfaced and correctly applied to a query in another domain.

  • Example: a preference for self-paced learning in education should transfer to recommending a self-guided tour in travel.
  • Score: fraction of CMRT items answered correctly.

• DCA: Measures whether a system can surface two simultaneously valid but conflicting preferences from different domains and produce a jointly optimizing response.

  • Example: a preference for nature trails and a sensitivity to prolonged sun exposure should lead to a recommendation such as an early-morning hike through shaded forests.
  • Score: fraction of DCA items answered correctly with both constraints preserved.

Retrieval Scoring

CrossMemBench evaluates retrieval separately from downstream task accuracy. Each memory provider returns both the context shown to the agent and a raw retrieval dump containing the retrieved memory_id values. CrossMemBench uses these IDs to determine whether the gold memories required for each item were retrieved.

• CMRT retrieval: counted as successful when the item's single gold memory is present in the retrieved memories.
• DCA retrieval: counted as successful only when both gold memories are present. Retrieving only one gold memory is treated as partial retrieval, while retrieving neither is treated as no retrieval.

This separation allows CrossMemBench to distinguish retrieval failures from cases where the required memories were available to the agent but were not used correctly.

For complete definitions and evaluation methodology, see our paper.

License

The evaluation code on our GitHub is released under the MIT License, while the CrossMemBench dataset on HuggingFace is released under the CC BY 4.0 License.

Citation

If you use CrossMemBench in your research, please cite:

@inproceedings{
  tiwari2026crossmembench,
  title={CrossMemBench: A Memory Benchmark for Cross-Domain Preference Transfer and Conflict Preservation},
  author={Devesh Tiwari},
  booktitle={COLM 2026 The 2nd Workshop on Lifelong Agents: Learning, Aligning, and Evolving},
  year={2026},
  url={https://openreview.net/forum?id=2BbJhtU7wa}
}