Moritze-Chenyu/BeyondRecall

Beyond Recall

A Fine-Grained Multimodal Benchmark for Long-Term Agent Memory

Evaluation Code | Download and Evaluation | Sources | License Scope | Citation

Beyond Recall evaluates retrieval, reasoning and management of information across long conversations with images, videos and documents. This release contains the complete current benchmark with native media and publication documentation.

Contents

benchmark/
  retrieval/              # Four task JSON files
  reasoning/              # Three task JSON files
  management/             # Four task JSON files
  images/                 # Images and release PDF files
  videos/                 # Original packaged video files, including audio tracks
  README.md               # Original release documentation
  RELEASE_STATS.json
  PDF_PREPROCESSING.json
  SHA256SUMS              # Original, frozen benchmark inventory
metadata/                 # Task statistics, media manifest, PDF sources and version
checksums/SHA256SUMS      # Full Hugging Face publication inventory
docs/CHANGELOG.md        # Data-version history

Paths inside task JSON are relative to benchmark/, not the repository root. Each task file is an array of conversation trajectories; one trajectory can have more than one question. IDs are local to each task. The Hub configuration names combine the dimension and task, and each exposes an evaluation-only test split. Only the 11 task JSON files are configured as data. Auxiliary metadata JSON is excluded from dataset configurations.

Statistics

Quantity	Count
Tasks	11
Conversation trajectories	1,163
Question-answer pairs	1,269
Timestamped sessions	23,775
Dialogue turns	345,738
Questions with explicit cutoff timestamps	40

Media type	Referenced packaged files	Exact-byte distinct content in the current release
Images	4,378	1,720
PDFs	93	73
Videos	338	314

The construction-reported source counts (1,865 images, 93 PDFs and 314 videos) are preserved in benchmark/RELEASE_STATS.json. They are not the current-file SHA256 deduplication counts shown above. File references and benchmark rows remain unchanged; repeated files are retained to preserve evaluation paths. The complete benchmark occupies approximately 7.14 GB before documentation artwork and metadata.

Tasks

Dimension	Task	Trajectories	QA pairs
Management	Conflict Detection	71	71
Management	Long Term Semantic Summary	71	71
Management	Memory Update	116	156
Management	Refusal To Answer	108	108
Reasoning	Causal Intent Inference	121	121
Reasoning	Cross Modal Relational Reasoning	73	124
Reasoning	Temporal Logic State Evolution	110	125
Retrieval	Cross Modal Semantic Matching	116	116
Retrieval	High Density Information Extraction	123	123
Retrieval	Property State Retrieval	139	139
Retrieval	Temporal Anchoring	115	115

Task Examples

The figure is from the accompanying paper. Click to view its original resolution.

Data Schema

Each record contains id, persons, conversation and questions. Sessions contain date and turns; turns contain person, content and optional media and timestamp. Questions contain question, answer and optional media and timestamp. Media paths refer to packaged files under benchmark/. The public reference answers are for evaluation; never supply them to the answering model. Preserve task-local IDs and question indices when matching predictions.

Identifiers and optional fields

Trajectory id values are numbered independently within each task. A unique question key is dimension/task/trajectory_id/qN, where N is the zero-based index in the trajectory's questions array. For example, management/memory_update/0/q0 identifies the first question in trajectory 0 of that task. Question indices are positions, not independent stored IDs.

media is a list of path strings relative to benchmark/. Treat an absent media, a loaded null, or an empty list as no attached media ([]). A present path must resolve to an actual file; missing files are input errors and must not be silently skipped. Treat an absent or null optional timestamp as unspecified (None). Empty or malformed non-null timestamps are errors, not permission to use unrestricted history. date is required for every session.

Time parsing and history visibility

Current session dates and question timestamps use English strings such as 8:27 am on 23 October, 2025, parsed with %I:%M %p on %d %B, %Y. English month names and AM/PM parsing require an English-compatible time locale. The evaluator additionally accepts seconds (%I:%M:%S %p on %d %B, %Y), 24-hour time (%H:%M on %d %B, %Y), %d %B %Y %H:%M, and ISO 8601 timestamps. ISO values with offsets are converted to UTC. Values without a timezone are assigned UTC as an evaluation normalization convention; the dataset does not establish the physical timezone of the depicted events. Do not substitute the machine's local timezone or rewrite the original strings.

For a question with a timestamp, a history turn is visible when its effective time is less than or equal to the question cutoff. A turn's explicit timestamp is used when present; otherwise it inherits its session's date. The current release has no explicit turn timestamps. Filtering is performed separately for each question before media preparation, caption input selection, memory write operations or retrieval-index construction. Future turns and their media must not enter that question's memory state. Question-attached media is query input, not a historical turn.

Retain only sessions with visible turns. Sort turns by effective time and their original index, and sessions by session date and original index. If an explicitly timestamped retained turn precedes its enclosing session date and the latter is after the cutoff, the evaluator uses the earliest retained turn time as that session's visible date. Without a question timestamp, use all supplied history in chronological order. Parsing failures stop evaluation rather than exposing full history. The GitHub evaluator starts an isolated memory state for each question after selecting its visible history.

Conversation text is English. Original media, speech, OCR quotes and identifiers can contain other languages. The original files are retained rather than translated or re-encoded merely to change language or reduce upload size.

Data Version

The data version is 1.0.0, recorded in metadata/VERSION.json. This identifies the evidence-preserving cropped-PDF benchmark; Hub revision IDs separately identify a repository snapshot, including documentation. Use the same pinned Hub revision for task records and native media; see Usage. Changes are recorded in Changelog.

PDF Revision

This is the evidence-preserving cropped-PDF release. Ninety-two PDF files retain all annotated evidence pages plus boundary pages, with optional other pages up to 10. Mandatory evidence can exceed 10 pages: 40 cropped files exceed that target. One PDF remains original because its source page annotations are invalid. In total, 41 current PDFs have more than 10 pages. Discarded pages are not provided as text. benchmark/PDF_PREPROCESSING.json preserves original page mappings and per-file hashes. Earlier full-PDF experiments are not directly comparable to this revised input.

Provenance and Intended Use

The paper describes human reconstruction of media-grounded memory tasks and injection into longer carrier conversations. Carrier text is adapted from LoCoMo and MMDU. Source-dataset annotations from MMLongBench-Doc guide fixed PDF page selection. Attribution and source-specific conditions are listed in Sources and Licensing.

Construction, annotation and review

Experts define capability boundaries and author seed tasks grounded in real media. Language models assist high-level storyline exploration. Generated memory drafts, questions, answers and evidence annotations are discarded after storyline abstraction; experts reconstruct the final fragments, timestamps, evidence relations, questions and reference answers from the source materials. Atomic tasks are translated into English and injected into long carrier conversations with deterministic, time-ordered scripts. Automatic checks and human review examine media grounding, temporal consistency, ambiguity, answer leakage and evidence sufficiency, including conflicts, updates and refusal cases. Private evidence annotations used during construction are not model inputs.

The packaged release uses dialogue speaker identifiers and anonymized publication metadata. This does not imply that names, faces, locations, speech or identifiers inside original media have been comprehensively redacted; native media is preserved.

This benchmark is intended for evaluating agent-memory behavior. Adapted carrier conversations, sampled media and question authoring choices can introduce source and topic biases; results should not be interpreted as representative of all users or cultures. Public answer-bearing test data should not be used as training data for a model whose benchmark performance is then reported as held-out evaluation.

Download and Verification

Download the entire repository with snapshot_download to obtain actual media bytes. load_dataset loads task records and path strings; it does not materialize all referenced images, PDFs or videos for this custom JSON schema.

# Run from the downloaded repository root.
sha256sum -c checksums/SHA256SUMS
(cd benchmark && sha256sum -c SHA256SUMS)

The original benchmark fingerprint remains 27f12c8e8f9b99a435dc7e1525eef0d00f44added1d0a04c612cd1bfbb2bbe23. See Usage for the GitHub evaluator and local loading examples.

License and Citation

Original-contribution licensing and public paper metadata are pending owner input. The top-level LICENSE records that status. Upstream data conditions and independent media rights are preserved; repository code licenses are not blanket media licenses. CITATION.cff and Citation provide a provisional anonymous manuscript entry without a fabricated DOI or arXiv URL.