Dataset Viewer

The dataset viewer should be available soon. Please retry later.

🧠 MemoryVCD · Full Edition

323 users · 12 domains · 535,698 interactions

📄 Paper · 📥 PDF · 💻 MemoryCD Code · 🔗 URL Edition · 🖼️ Full Edition

Real user histories for evaluating long-term, cross-domain personalization. MemoryVCD extends the MemoryCD data release with product image references, complete metadata for the saved candidate pools, and explicit evaluation inputs. Full adds the available cached product images.

📄 Paper and release relationship

MemoryCD: Benchmarking Long-Context User Memory of LLM Agents for Lifelong Cross-Domain Personalization — Zhang et al., Findings of EMNLP 2026.

The paper studies memory grounded in real Amazon review histories, with four tasks: rating prediction, item ranking, review summarization, and review generation. It evaluates single-domain and cross-domain memory. See the paper for methods and results.

This repository is the MemoryVCD extended release, not a claim that every added visual feature or candidate file was evaluated in the MemoryCD paper. Statistics below describe the released files.

Repository Purpose Product images
MemoryCD Original release; unchanged Original item metadata
MemoryVCD Extended histories, candidate metadata and evaluation inputs URL strings
MemoryVCD-Full Same data plus existing image cache 243,394 embedded image records

📊 At a glance

Statistic Count
Users 323
Domains 12: 4 evaluation targets + 8 memory domains
Interaction records 535,698
History product/domain pairs 276,418
All product/domain metadata records 481,333
Unique sampled negative product/domain pairs 211,364
Negative items outside the history product set 204,915
Missing metadata for saved negatives 0
Product image URLs 456,761
Valid cached image records (Full) 243,394

Counts use (domain, parent_asin) as the item key. A negative for one user can be a historical item for another; history and negative counts therefore overlap.

🗂️ Browse the data

Subset Rows Contents
interactions 535,698 One review per row
products 481,333 History and candidate item metadata
negative_items 211,364 Metadata for all saved sampled negatives
users 323 User index and activity counts
evaluation_windows 15,504 Memory boundaries and test targets
evaluation_candidates 19,751 Positive and negative item IDs with source provenance
product_images 243,394 Cached product images with native image fields

All subsets use test to label the evaluation corpus. This is not a random train/test split: evaluation_windows defines memory and test targets within each history.

🎯 Negative items and sampling

products includes both historical items and all items required by the saved ranking/choice candidates. negative_items is a convenient filtered view of those same metadata records; it does not add new interactions or change the candidate pools.

Domain Unique negative items Outside history Missing metadata
Beauty_and_Personal_Care 51,966 49,416 0
Books 53,433 52,571 0
Electronics 52,641 51,075 0
Home_and_Kitchen 53,324 51,853 0
Total 211,364 204,915 0
Candidate task Saved rows Negatives per row Source
Item ranking 9,876 19 Saved random candidate pools
Item preference choice 9,875 3 Saved hard-negative pools

No resampling was performed. Candidate order and IDs are preserved. The saved hard-negative files contain 3 negatives; any description expecting 4 does not match these files. Base and time-cutoff sources are distinct: filter task, source_file, and cutoff before evaluation. Time-axis source files cover the original first 100 users, while evaluation windows cover all 323 users. Missing candidate rows are not fabricated.

Join evaluation_candidates.neg_asins to negative_items.parent_asin within the same domain. candidate_reference_count counts occurrences across all saved candidate rows; candidate_tasks lists the tasks using each item. The coverage audit is in reports/negative_metadata_coverage.json.

🚀 Load and join

from datasets import load_dataset

repo = "WZDavid/MemoryVCD-Full"
reviews = load_dataset(repo, "interactions", split="test")
candidates = load_dataset(repo, "evaluation_candidates", split="test")
negatives = load_dataset(repo, "negative_items", split="test")

meta = {(r["domain"], r["parent_asin"]): r for r in negatives}
example = candidates[0]  # Select your intended task/source/cutoff for evaluation.
negative_metadata = [meta[(example["domain"], a)] for a in example["neg_asins"]]
print([(r["parent_asin"], r["title"]) for r in negative_metadata])

For positive item metadata, use the products subset. To read cached images:

images = load_dataset("WZDavid/MemoryVCD-Full", "product_images",
                      split="test", streaming=True)
example_image = next(iter(images))
print(example_image["domain"], example_image["parent_asin"])
image = example_image["image"]  # PIL image; cached bytes are embedded in Parquet.

🌍 Domain statistics

The first four domains are evaluation targets (314,116 interactions). The remaining eight provide additional memory (221,582 interactions). Users overlap across domains.

Domain Users Interactions History products
Beauty_and_Personal_Care 323 76,094 33,475
Books 323 79,874 48,054
Electronics 323 44,268 25,441
Home_and_Kitchen 323 113,880 60,900
All_Beauty 293 2,438 1,336
Arts_Crafts_and_Sewing 319 18,588 10,508
Automotive 321 6,638 4,348
Grocery_and_Gourmet_Food 323 48,960 19,154
Health_and_Household 323 80,391 35,543
Movies_and_TV 278 9,395 8,433
Office_Products 323 28,325 14,032
Sports_and_Outdoors 323 26,847 15,194
Total (unique users) 323 535,698 276,418

🖼️ Images and metadata

Domain Required metadata Image URLs Cached image records
Beauty_and_Personal_Care 82,891 82,890 49,225
Books 100,625 78,730 49,539
Electronics 76,516 76,502 47,696
Home_and_Kitchen 112,753 112,753 72,262
All_Beauty 1,336 1,336 781
Arts_Crafts_and_Sewing 10,508 10,499 2,229
Automotive 4,348 4,348 1,397
Grocery_and_Gourmet_Food 19,154 19,149 2,911
Health_and_Household 35,543 35,543 7,689
Movies_and_TV 8,433 5,840 244
Office_Products 14,032 13,977 3,936
Sports_and_Outdoors 15,194 15,194 5,485
Total 481,333 456,761 243,394

Full includes all matched valid images in the release cache, not 100% image coverage: 50.57% of required metadata rows and 30.23% of history products have cached images. One invalid cached image was excluded. External URLs may expire. Product photos are separate from user-uploaded review photos; review image bytes are not bundled.

🧾 Key fields and files

Table / file Important fields or contents
interactions user_id, domain, parent_asin, rating, timestamp, review title/text, image_url
products / negative_items title, description, category, store, average rating, image URL; in_user_history, cache availability
evaluation_candidates task, source_file, cutoff, pos_asin, neg_asins
evaluation_windows Target indices, memory prefix/boundary, source domains
product_images (Full) (domain, parent_asin), embedded image, source URL and image hash
users/users_12domains.jsonl.gz Original nested histories, one user per line
metadata/*.jsonl.gz Per-domain metadata including required candidate items
evaluation/source_candidates/ Original saved hard-negative files
tools/release_dataset_helpers.py Load nested histories and reconstruct evaluation inputs
reports/, manifest.json Coverage checks, validation, file sizes and SHA-256 hashes

Metadata uses the evaluator's slim fields: title up to 200 characters and description up to 400; price is not included. image_shard paths refer to the Full repository. Join images with (domain, parent_asin). The users subset is an index; use the compressed histories for the complete nested records.

⚙️ Evaluation and limitations

Setting Released definition
Time axis Five cutoffs: 0, .25, .5, .75, 1; a three-interaction target window
Domain axis D0–D5 and Dtx; final three target-domain interactions
Chronology Memory precedes the target boundary; see helper functions for exact reconstruction
Ordering source_user_index, source_interaction_index and stable interaction IDs preserve provenance
Repeated keys 10,404 excess records share user/domain/item/timestamp; retained to preserve timelines
Selection bias Users have substantial cross-domain activity; not a representative random sample
Review images Extra-eight-domain extraction did not retain review-image fields; absence is not proof of no original image

The same histories/windows support multiple tasks; task counts must not be summed as independent users. Exact source ordering and candidate versions matter for reproducibility.

📚 Citation and source

Please cite the foundational MemoryCD paper and identify the MemoryVCD edition/revision used:

@inproceedings{zhang-etal-2026-memorycd,
  title = "{M}emory{CD}: Benchmarking Long-Context User Memory of {LLM} Agents for Lifelong Cross-Domain Personalization",
  author = "Zhang, Weizhi and Wei, Xiaokai and Huang, Wei-Chieh and Hui, Zheng and Wang, Chen and Gong, Michelle and Yu, Philip S.",
  booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2026",
  year = "2026",
  pages = "38554--38575",
  url = "https://aclanthology.org/2026.findings-emnlp.1804/",
  doi = "10.18653/2026.findings-emnlp.1804"
}

Derived from Amazon Reviews 2023. Original reviews and images retain applicable source terms and third-party rights. This packaging does not grant new rights to those materials.

Data release: 2026-11-12 · Negative-item view and documentation: 2026-11-18. Shared tables are identical across URL and Full editions; Full additionally includes cached image bytes.

Downloads last month
117