LongMemEval Dataset

July 24, 2026 · View on GitHub

User Memory Evaluation guide

LongMemEval is a comprehensive benchmark for evaluating long-term interactive memory in chat assistants, published by UCSB et al. It contains 500 carefully designed questions across six categories, covering information extraction, cross-session reasoning, temporal reasoning, knowledge update, preferences, and abstention-style behavior.

Files

FileSizeDescription
longmemeval_oracle.json~14.7 MBOracle version with evidence sessions only
longmemeval_s_cleaned.json~264.5 MBLongMemEval-S: ~48 sessions / ~115K tokens per question
longmemeval_m_cleaned.json~2.6 GBLongMemEval-M: ~500 sessions / ~1.5M tokens per question

Only README.md and prepare_longmemeval.py are intended to be version-controlled by OmniMemEval. Downloaded JSON files are generated artifacts and should remain local.

Dataset Statistics

MetricValue
Total questions500
S-version context~115K tokens / question (~48 sessions)
M-version context~1.5M tokens / question (~500 sessions)
Question categories6

Question Categories

CategoryCountDescription
temporal-reasoning133Requires understanding temporal order and time-related information
multi-session133Requires integrating information across multiple sessions
knowledge-update78Information changes over time; must answer with the latest version
single-session-user70Extract user-related facts from a single session
single-session-assistant56Extract assistant-related facts from a single session
single-session-preference30Extract user preferences from a single session

Data Structure

Top-level JSON array of 500 question objects. Each question contains:

FieldTypeDescription
question_idstringUnique identifier, e.g., "gpt4_2655b836"
question_typestringOne of the 6 categories above
questionstringQuestion text
answerstringGold answer
question_datestringTimestamp when the question is asked, e.g., "2023/04/10 (Mon) 23:07"
haystack_dateslistDates of haystack sessions
haystack_session_idslistIDs of haystack sessions
haystack_sessionslistFull session dialogue content
answer_session_idslistSession IDs containing evidence for the answer

The S and M versions share the same structure; M includes significantly more filler sessions.

Data Preparation

The JSON files are downloaded directly from Hugging Face xiaowu0162/longmemeval-cleaned. The default download is the S variant, matching OmniMemEval's runner default.

python data/longmemeval/prepare_longmemeval.py                       # download S variant (~265 MB, used by default)
python data/longmemeval/prepare_longmemeval.py --variant oracle s m  # download all variants (M is ~2.6 GB)
python data/longmemeval/prepare_longmemeval.py --force               # overwrite existing files

The script validates each downloaded JSON file before replacing the local copy.

License And Redistribution

xiaowu0162/longmemeval-cleaned is published on Hugging Face with the MIT license. The OmniMemEval code license does not override the upstream LongMemEval dataset license.

Evaluation

  • Metric: LLM-as-a-Judge Accuracy (default judge: gpt-4o-mini)
  • Pipeline: Ingest sessions → Search relevant context → Generate answer → LLM judge scores correctness
  • 6 question categories: Single-session user facts, assistant facts, preferences, multi-session reasoning, temporal reasoning, and knowledge update

References